Search

While the influence of surface melt on Antarctic ice shelf stability can be large, the duration and affected area of melt events are often small. Therefore, melt events are difficult to capture with remote sensing, as satellite sensors always face the trade-off between spatial and temporal resolution. To overcome this limitation, we developed UMelt: a surface melt record for all Antarctic ice shelves with a high spatial (500 m) and high temporal (12 h) resolution for the period 2016–2021. Our approach is based on a deep learning model, specifically a U-Net, which was developed in Google Earth Engine. The U-Net combines microwave remote sensing observations from three sources: Sentinel-1, Special Sensor Microwave Imager/Sounder (SSMIS), and Advanced Scatterometer (ASCAT). The U-Net was trained on the Shackleton Ice Shelf for melt seasons 2017–2021, using the fine-scale melt patterns of Sentinel-1 as reference data and SSMIS, ASCAT, a digital elevation model, and multi-year Sentinel-1 melt fraction as predictors. The trained U-Net performed well on the Shackelton Ice Shelf for test melt season 2016–2017 (accuracy: 91.3%; F1-score: 86.9%), and the Larsen C Ice Shelf, which was not considered during training (accuracy: 91.0%; F1-score: 89.3%). Using the trained U-Net model, we have successfully developed the UMelt record. UMelt allows Antarctic-wide surface melt to be detected at a small scale while preserving a high temporal resolution, which could lead to new insights into the response of ice shelves to a changing atmospheric forcing.

Deep learning-based models have revolutionized biomedical image and video segmentation, enabling precise and automated analysis of complex structures. This advancement is particularly critical in the study of zebrafish cardiovascular videos, where accurate segmentation of the heart is essential for understanding cardiac function and development. In this work, we focus on the Zebrafish Automated Cardiac Analysis Framework (ZACAF), utilizing the U-net architecture to achieve high-accuracy segmentation of the zebrafish heart from video frames to calculate important cardiovascular parameters.After multiple collaborations with researchers studying different phenotypes in zf using ZACAF, to enhance the generalizability of our model across varying datasets and conditions, we employed transfer learning techniques, leveraging pre-trained models to adapt to new data efficiently. Additionally, we incorporated Test-Time Augmentation (TTA) to further improve model robustness and accuracy by applying various transformations to the input data during inference. This approach proposed a systematic solution for adopting ZACAF in broader genetic studies using deep learning algorithms. Recognizing the importance of temporal dynamics in video data, we extended our work to integrate temporal features into the segmentation model. By analyzing changes between consecutive frames, we aim to capture the heartbeat dynamics more effectively, providing a comprehensive tool for cardiac analysis in zebrafish embryos. Our approach not only advances the field of biomedical video segmentation but also contributes to the broader understanding of cardiac function in developmental biology.

La flora marina, en especial las praderas marinas y praderas de algas, posee un rol esencial en la sostenibilidad de los ecosistemas costeros. Por esto, existe una elevada necesidad en desarrollar herramientas que permitan monitorizar de forma precisa la flora marina. La teledetección por satélite es una tecnología eficiente y económica para poder elaborar mapas de hábitats bentónicos o de tipos de fondos, permitiendo caracterizar los fondos a lo largo del tiempo. Para la generación de estos mapas, en la actualidad típicamente se implementan técnicas basadas en píxel, como modelos Machine Learning convencionales, y otras técnicas basadas en objetos, como el OBIA. No obstante, el avance de las técnicas Deep Learning permite entrenar modelos que aprovechen mejor las características de los fondos marinos para la generación de mapas de elevada precisión. En este trabajo se expone el uso del modelo U-Net, basado en Redes Neuronales Convoluciones (CNN), para la generación de mapas de hábitats bentónicos a partir de imágenes satelitales multiespectrales. La zona de estudio es El Río, estrecho entre las islas de La Graciosa y Lanzarote, islas Canarias, donde existe una elevada biodiversidad con la presencia de multitud de especies como praderas de Cymodocea nodosa, Caulerpa prolifera y algas rojas filamentosas, entre otras especies., Marine flora, especially seagrass and seaweed meadows, plays an essential role in the sustainability of marine ecosystems. Therefore, there is a great need to develop tools to accurately monitor the marine flora. Satellite remote sensing is an efficient and economical technology for mapping benthic habitats or bottom types, allowing the characterization of the seabed over time. For the generation of these maps, techniques pixel-based techniques, such as conventional Machine Learning models, and object-based techniques, such as OBIA, are currently typically implemented. However, the advancement of Deep Learning techniques allows training models that take better advantage of seabed features for the generation of high accurate maps. In this paper we present the use of the U-Net model, based on Convolutional Neural Networks (CNN), for the generation of benthic habitat maps based on multispectral satellite imagery. The study area is El Río, a strait between the islands of La Graciosa and Lanzarote, Canary Islands, where there is a high biodiversity with the presence of a multitude of species such as Cymodocea nodosa meadows, Caulerpa prolifera and red filamentous algae, among other species., Esta investigación ha sido parcialmente subvencionada por el Ministerio para la Transición Ecológica y el Reto Demográfico mediante el proyecto TARA (Ref. SPIP2022-02897)., Peer Reviewed, Postprint (published version)

Cloud is an aerosol consisting of visible mass of miniature liquid droplets, frozen crystal, or other particles suspended in the atmosphere. The study of atmospheric clouds is crucial for us to better understand and predict the behaviors of clouds, which has implications for climate, weather, aviation safety, agriculture, and energy production. Convolutional neural network (CNN) method is applied to train an atmospheric cloud image detection model to identify the presence of cloud and classify them. Supervised learning method is applied to train the model such that the machine is given labeled cloud image dataset to learn how to classify and predict the presence of cloud. U-Net architecture is used to train the atmospheric cloud image detection model because the architecture has the highest performance in image segmentation especially object detection in satellite images. The 38-Cloud Dataset which is used to train the model, is obtained from Landsat 8 Earth observation satellite. The dataset is randomly divided into training set (75% of the total images) and validation set (25% of the total images). Following this, the dataset is preprocessed and transformed into tensors to train the model. The training has been carried out for 50 epochs. Apart from the U-Net architecture proposed, the architecture is further modified with ResNet34 and VGG16 and the performance of each model is studied. The recognition accuracy obtained for atmospheric cloud image detection trained with the dataset achieved 97%. With this accuracy, U-Net architecture can be justified as a powerful and suitable convolutional neural network in performing atmospheric cloud image detection.&nbsp

Accurate segmentation of river water in close-range Remote Sensing (RS) images is vital for efficient environmental monitoring and management. However, this task poses significant difficulties due to the dynamic nature of water, which exhibits varying colors and textures reflecting the sky and surrounding structures along the riverbanks. This study addresses these complexities by evaluating and comparing several well-known deep-learning (DL) techniques on four river scene datasets. To achieve this, we fine-tuned the recently introduced 'Segment Anything Model' (SAM) along with popular DL segmentation models such as U-Net, DeepLabV3+, LinkNet, PSPNet, and PAN, all using ResNet50 pre-trained on ImageNet as a backbone. Experimental results highlight the diverse performances of these models in river water segmentation. Notably, fine-tuned SAM demonstrates superior performance, followed by U-Net(ResNet50), despite their higher computational costs. In contrast, PSPNet(ResNet50), while less effective, proves to be the most efficient in terms of execution time. In addition to these findings, we introduce a novel river water segmentation dataset, LuFI-RiverSnap. v1 (Dataset link), characterized by a more diverse range of scenes and accurate masks compared to existing datasets. To facilitate reproducible research in remote sensing and computer vision, we release the implementations of the fine-tuned SAM model (Code link). The findings from this research, coupled with the presented dataset and the accuracy achieved by fine-tuned SAM segmentation, can support tracking river changes, understanding river water level trends, and exploring river ecosystem dynamics. These can also provide valuable insights for practitioners and researchers seeking models tailored to specific image characteristics with practical means in disaster risk reduction, such as rapid assessments of inundations during floods or automatic extractions of gauge data in watersheds.

Eliminating Multi-Path Interference (MPI) stands as a significant unresolved challenge in the domain of depth estimation using Time-of-Flight (ToF) cameras. ToF data is typically influenced by significant noise and artifacts stemming from MPI. Although a variety of conventional methods have been suggested to enhance ToF data quality, the application of machine learning techniques has been infrequent, primarily due to the scarcity of authentic training data with accurate depth information. This paper introduces an approach that eliminates the dependency on labeled real-world data within the learning framework. We employ a U-Net trained on the data with ground truth in a supervised manner, enabling it to leverage multi-frequency ToF data for MPI correction. Concurrently, we compare three channels as input with one channel and two channels. Our experimental results convincingly showcase the effectiveness of this approach in reducing noise in real-world data.

Accurate medical image segmentation is crucial for safely and effectively administering radiation therapy in cancer treatment. State of the art methods for automatic segmentation of 3D images are currently based on the U-net machine learning architecture. The current U-net models are large, often containing millions of parameters. However, the size of these machine learning models can be decreased by removing parts of the models, in what is called pruning. One algorithm, called simultaneous training and pruning (STAMP) has shown capable of reducing the model sizes upwards of 80% while keeping similar or higher levels of performance for medical image segmentation tasks. This thesis investigates the impact of using the STAMP algorithm to reduce model size and inference time for medical image segmentation on 3D images, using one MRI and two CT datasets. Surprisingly, we show that pruning convolutional filters randomly achieves performance comparable, if not better than STAMP, provided that the filters are always removed from the largest parts of the U-net. Inspired by these results, a modified "Flat U-net" is proposed, where an equal number of convolutional filters are used in all parts of the U-net, similar to what was achieved after pruning with our simplified pruning algorithm. The modified U-net achieves similar levels of test dice score as both a regular U-net and the STAMP pruning algorithm, on multiple datasets while avoiding pruning altogether. In addition to this the proposed modification reduces the model size by more than a factor of 12, and the number of computations by around 35%, compared to a normal U-net with the same number of input-layer convolutional filters.

This paper addresses the critical need for advanced real-time vehicle detection methodologies in Vehicle Intelligence Systems (VIS), especially in the context of using Unmanned Aerial Vehicles (UAVs) for data acquisition in severe weather conditions, such as heavy snowfall typical of the Nordic region. Traditional vehicle detection techniques, which often rely on custom-engineered features and deterministic algorithms, fall short in adapting to diverse environmental challenges, leading to a demand for more precise and sophisticated methods. The limitations of current architectures, particularly when deployed in real-time on edge devices with restricted computational capabilities, are highlighted as significant hurdles in the development of efficient vehicle detection systems. To bridge this gap, our research focuses on the formulation of an innovative approach that combines the fractional B-spline wavelet transform with a tailored U-Net architecture, operational on a Raspberry Pi 4. This method aims to enhance vehicle detection and localization by leveraging the unique attributes of the NVD dataset, which comprises drone-captured imagery under the harsh winter conditions of northern Sweden. The dataset, featuring 8450 annotated frames with 26,313 vehicles, serves as the foundation for evaluating the proposed technique. The comparative analysis of the proposed method against state-of-the-art detectors, such as YOLO and Faster RCNN, in both accuracy and efficiency on constrained devices, emphasizes the capability of our method to balance the trade-off between speed and accuracy, thereby broadening its utility across various domains., Validerad;2024;Nivå 2;2024-09-03 (joosat);Full text: CC BY License

This paper addresses the critical need for advanced real-time vehicle detection methodologies in Vehicle Intelligence Systems (VIS), especially in the context of using Unmanned Aerial Vehicles (UAVs) for data acquisition in severe weather conditions, such as heavy snowfall typical of the Nordic region. Traditional vehicle detection techniques, which often rely on custom-engineered features and deterministic algorithms, fall short in adapting to diverse environmental challenges, leading to a demand for more precise and sophisticated methods. The limitations of current architectures, particularly when deployed in real-time on edge devices with restricted computational capabilities, are highlighted as significant hurdles in the development of efficient vehicle detection systems. To bridge this gap, our research focuses on the formulation of an innovative approach that combines the fractional B-spline wavelet transform with a tailored U-Net architecture, operational on a Raspberry Pi 4. This method aims to enhance vehicle detection and localization by leveraging the unique attributes of the NVD dataset, which comprises drone-captured imagery under the harsh winter conditions of northern Sweden. The dataset, featuring 8450 annotated frames with 26,313 vehicles, serves as the foundation for evaluating the proposed technique. The comparative analysis of the proposed method against state-of-the-art detectors, such as YOLO and Faster RCNN, in both accuracy and efficiency on constrained devices, emphasizes the capability of our method to balance the trade-off between speed and accuracy, thereby broadening its utility across various domains., Validerad;2024;Nivå 2;2024-09-03 (joosat);Full text: CC BY License

Brain tumor segmentation, among various tasks in medical image analysis, has garnered significant attention in the research community. Despite continuous efforts by researchers, accurate brain tumor segmentation remains a key challenge. This challenge arises due to various factors, including location uncertainty, morphological uncertainty, low contrast imaging, annotation bias, and data imbalance. Magnetic resonance imaging (MRI) plays a vital role in providing detailed images of the brain, enabling the extraction of crucial information about the tumor’s shape, size, and location. In literature, deep learning algorithms have shown their efficiency in dealing with semantic segmentation, particularly the U-Net architecture. The latter has demonstrated impressive performance in Medical image segmentation. In this paper, a U-Net-based architecture for brain tumor segmentation is proposed. To further enhance the segmentation performance of our model, a novel ensemble learning method, EnsUNet, is introduced by integrating four pre-trained networks namely MobileNet, DeepLabV3+, ResNet, and DenseNet as the encoder within the U-Net architecture. The conducted experimental evaluation demonstrates promising results, achieving an Intersection over Union (IoU) score of 0.86, a Dice Coefficient (DC) of 0.92, and an accuracy of approximately 0.99. These findings underscore the effectiveness of our proposed EnsUNet for accurately segmenting brain tumors. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.

A convolutional neural network (CNN) was used to construct a semantic segmentation model to examine the abundance of Phycosiphon incertum by identifying the trace fossil regions in core section images. The abundance of trace fossils provides information about the past activities of benthic animals affected by paleoenvironmental conditions. To quantify the intensity of bioturbation, it is necessary to extract regions of trace fossils and measure the proportion of bioturbated and the observed area of the outcrop section. In this study, a U-Net-type CNN model was used with residual connections and attention mechanisms to identify the trace fossil Phycosiphon. The model was trained to recognize the relationships between core section images from the International Ocean Discovery Program Expedition 362 Site U1480 and manually annotated trace fossil images. After training, the model successfully classified the pixels of the background, outcrop, and Phycosiphon for core section images other than the training data set. The bioturbation intensity estimated from the image predicted by the model was nearly equal to that from the ground truth image. A long-term (approximately past 10 Myr) variation in Phycosiphon abundance was estimated by applying the model to the core section images at Site U1480. Phycosiphon abundance negatively correlated with the number of sandstone layer intercalations, but it was not affected by the sediment accumulation rates. These findings may reflect resistance of Phycosiphon producers to environmental stress. The model developed in this study can be used for other ichnotaxa to reveal the general tendency of variation in bioturbation intensity and ichnodiversity.

Accurate medical image segmentation is crucial for safely and effectively administering radiation therapy in cancer treatment. State of the art methods for automatic segmentation of 3D images are currently based on the U-net machine learning architecture. The current U-net models are large, often containing millions of parameters. However, the size of these machine learning models can be decreased by removing parts of the models, in what is called pruning. One algorithm, called simultaneous training and pruning (STAMP) has shown capable of reducing the model sizes upwards of 80% while keeping similar or higher levels of performance for medical image segmentation tasks. This thesis investigates the impact of using the STAMP algorithm to reduce model size and inference time for medical image segmentation on 3D images, using one MRI and two CT datasets. Surprisingly, we show that pruning convolutional filters randomly achieves performance comparable, if not better than STAMP, provided that the filters are always removed from the largest parts of the U-net. Inspired by these results, a modified "Flat U-net" is proposed, where an equal number of convolutional filters are used in all parts of the U-net, similar to what was achieved after pruning with our simplified pruning algorithm. The modified U-net achieves similar levels of test dice score as both a regular U-net and the STAMP pruning algorithm, on multiple datasets while avoiding pruning altogether. In addition to this the proposed modification reduces the model size by more than a factor of 12, and the number of computations by around 35%, compared to a normal U-net with the same number of input-layer convolutional filters.

Solcellspaneler har fått ökad uppmärksamhet som en betydande källa till förnybar energi på grund av den ökande medvetenheten om klimatförändringar och behovet av hållbara energilösningar. I Sverige har detta lett till en kraftig ökning av solcellsanläggningar. I och med ökningen av solcellspaneler i urbana miljöer, har behovet av att kartlägga och inventera dessa anläggningar växt. Framsteg inom artificiell intelligens (AI) och bildanalys har öppnat möjligheter för automatiserade metoder som effektivt kan identifiera och segmentera solcellspaneler. Syftet med detta arbete är att utforska potentialen hos traditionella AI-modeller, som Support Vector Machines (SVM) och Random Forest (RF), samt faltningsnätverk av typen U-net, för att identifiera och segmentera solcellspaneler i svenska flygfotodata. Vidare undersöks i arbetet hur dessa modeller reagerar på reducerad datamängd samt vad för sorts features som höjer de traditionella modellernas prestanda i syfte att inventera solcellspaneler. Under arbetet skapas ett dataset om 2268, 1152 RGBI-data, där solcellspaneler utgör 22,8 procent av pixlarna. Data är hämtad från Lantmäteriet och har en spatial upplösning på 0.16m/pixel. Tre modeller implementeras och jämförs under olika förhållanden. Flertalet features utvunna från datasetet presenteras och förändringar av prestanda vid träning med dessa features mäts. För en utvärdering av modellernas precision tränas de först på 70% av det totala datasetet och utvärderas på de resterande 30%. En andra utvärdering utförs med reducerad datamängd där 35% av den totala datamängden används för träning och 30% för utvärdering. Prestandamätningar utförs på samma dataset för alla modeller där traditionella modeller tränas på RGB, RGBI, RGBI+features. U-net-modellen tränas på RGB-data. Resultaten visar att U-net-modellen presterar bäst i syfte att segmentera solcellspaneler med en F1-score på 0.91 och MCC på 0.89. Näst bäst är RF med en F1-score på 0.81 samt MCC på 0.76. Vid halvering a, Solar panels have gained increased attention as a significant source of renewable energy due to the growing awareness of climate change and the need for sustainable energy solutions. In Sweden, this has led to a rise in solar installations. With the increase of solar panels in urban areas, the need for accurate mapping and inventory of these installations has grown. Advances in artificial intelligence (AI) and image analysis have opened possibilities for automated methods for high precision identification av segmentation of solar panels. These automated methods reduce time consumption, resource use and the risk of human error. The aim of this work is to explore the potential of traditional AI models such as Support Vector Machines (SVM) and Random Forest (RF), as well as convolutional neural networks with the U-net architecture, to identify and segment solar panels in Swedish aerial imagery. Furthermore, the study investigates how these models perform with reduced data quantities. The study also examines which types of features enhance the performance of traditional models for the purpose of inventorying solar panels. During the study, a dataset of 2268, 1152 RGBI data was created, where solar panels constitute 22.8 percent of the pixels. This data, sourced from Lantmäteriet, has a spatial resolution of 0.16m/pixel. Three models were implemented and compared under various conditions. Multiple features extracted from the dataset were presented and performance changes during training with these features were measured. For evaluation the models were first trained on 70% of the total dataset and evaluated on the remaining 30%. A second evaluation was conducted with reduced data, using 35% for training and 30% for evaluation. Performance measurements were carried out on the same dataset for all models, where the traditional models were trained on RGB, RGBI, and RGBI + features, while the U-net model was trained on RGB data. In evaluation the U-net model achieved the high

With the evolution of GPU performance, the interest of using AI for all kinds of purposes has risen. Companies today put a great amount of resources to find new ways of using AI to increase the value of their products or automating processes. An area in the wood industry where AI is widely used and studied is in defect detection. In this thesis, the combination of using AI and hyperspectral images is studied and evaluated in the case of segmenting defects in hardwood with a U- Net network structure. The performance is compared to another known method usually used when dealing with high-dimensional data: PLS-DA. This thesis also compares the use of RGB image data in combination with AI, to further analyze the usefulness that the hyperspectral data provide. The results showed signs of improvement when using hyperspectral images com- pared to RGB images when detecting blue stain and red heartwood defects. De- tection of the defects rot and knots did however show no sign of improvements. Due to the annotations being more accurate in the RGB data, the results from the hyperspectral data-fed networks would suggest that blue stain and red heartwood could be of interest regarding further investigation. Computational performance is shown to vary across the different reduction meth- ods, and the results from this thesis provides some insight that might aid in the reasoning regarding how to choose an appropriate reduction method.

Solcellspaneler har fått ökad uppmärksamhet som en betydande källa till förnybar energi på grund av den ökande medvetenheten om klimatförändringar och behovet av hållbara energilösningar. I Sverige har detta lett till en kraftig ökning av solcellsanläggningar. I och med ökningen av solcellspaneler i urbana miljöer, har behovet av att kartlägga och inventera dessa anläggningar växt. Framsteg inom artificiell intelligens (AI) och bildanalys har öppnat möjligheter för automatiserade metoder som effektivt kan identifiera och segmentera solcellspaneler. Syftet med detta arbete är att utforska potentialen hos traditionella AI-modeller, som Support Vector Machines (SVM) och Random Forest (RF), samt faltningsnätverk av typen U-net, för att identifiera och segmentera solcellspaneler i svenska flygfotodata. Vidare undersöks i arbetet hur dessa modeller reagerar på reducerad datamängd samt vad för sorts features som höjer de traditionella modellernas prestanda i syfte att inventera solcellspaneler. Under arbetet skapas ett dataset om 2268, 1152 RGBI-data, där solcellspaneler utgör 22,8 procent av pixlarna. Data är hämtad från Lantmäteriet och har en spatial upplösning på 0.16m/pixel. Tre modeller implementeras och jämförs under olika förhållanden. Flertalet features utvunna från datasetet presenteras och förändringar av prestanda vid träning med dessa features mäts. För en utvärdering av modellernas precision tränas de först på 70% av det totala datasetet och utvärderas på de resterande 30%. En andra utvärdering utförs med reducerad datamängd där 35% av den totala datamängden används för träning och 30% för utvärdering. Prestandamätningar utförs på samma dataset för alla modeller där traditionella modeller tränas på RGB, RGBI, RGBI+features. U-net-modellen tränas på RGB-data. Resultaten visar att U-net-modellen presterar bäst i syfte att segmentera solcellspaneler med en F1-score på 0.91 och MCC på 0.89. Näst bäst är RF med en F1-score på 0.81 samt MCC på 0.76. Vid halvering a, Solar panels have gained increased attention as a significant source of renewable energy due to the growing awareness of climate change and the need for sustainable energy solutions. In Sweden, this has led to a rise in solar installations. With the increase of solar panels in urban areas, the need for accurate mapping and inventory of these installations has grown. Advances in artificial intelligence (AI) and image analysis have opened possibilities for automated methods for high precision identification av segmentation of solar panels. These automated methods reduce time consumption, resource use and the risk of human error. The aim of this work is to explore the potential of traditional AI models such as Support Vector Machines (SVM) and Random Forest (RF), as well as convolutional neural networks with the U-net architecture, to identify and segment solar panels in Swedish aerial imagery. Furthermore, the study investigates how these models perform with reduced data quantities. The study also examines which types of features enhance the performance of traditional models for the purpose of inventorying solar panels. During the study, a dataset of 2268, 1152 RGBI data was created, where solar panels constitute 22.8 percent of the pixels. This data, sourced from Lantmäteriet, has a spatial resolution of 0.16m/pixel. Three models were implemented and compared under various conditions. Multiple features extracted from the dataset were presented and performance changes during training with these features were measured. For evaluation the models were first trained on 70% of the total dataset and evaluated on the remaining 30%. A second evaluation was conducted with reduced data, using 35% for training and 30% for evaluation. Performance measurements were carried out on the same dataset for all models, where the traditional models were trained on RGB, RGBI, and RGBI + features, while the U-net model was trained on RGB data. In evaluation the U-net model achieved the high

With the evolution of GPU performance, the interest of using AI for all kinds of purposes has risen. Companies today put a great amount of resources to find new ways of using AI to increase the value of their products or automating processes. An area in the wood industry where AI is widely used and studied is in defect detection. In this thesis, the combination of using AI and hyperspectral images is studied and evaluated in the case of segmenting defects in hardwood with a U- Net network structure. The performance is compared to another known method usually used when dealing with high-dimensional data: PLS-DA. This thesis also compares the use of RGB image data in combination with AI, to further analyze the usefulness that the hyperspectral data provide. The results showed signs of improvement when using hyperspectral images com- pared to RGB images when detecting blue stain and red heartwood defects. De- tection of the defects rot and knots did however show no sign of improvements. Due to the annotations being more accurate in the RGB data, the results from the hyperspectral data-fed networks would suggest that blue stain and red heartwood could be of interest regarding further investigation. Computational performance is shown to vary across the different reduction meth- ods, and the results from this thesis provides some insight that might aid in the reasoning regarding how to choose an appropriate reduction method.

Time-difference-of-arrival (TDOA) estimation from GCC-PHAT is not always as straight forward as finding the maximum peak. This work views the GCC output as an image, with time on the vertical axis and TDOA horizontally, to explore if image-to-image machine learning methods can make a more robust filter. The Structure from Sound Database provides audio recorded with a distributed microphone setup and a moving sound source. The audio was fed to GCC-PHAT without pre-processing, and images were produced for batch processing. The ground truth, the direct-path TDOA, shows a continuous curve through time. The GCC output image has a similar curve, but obscured by noise and not at all times texturally different from the multi-path components. The main approach tested is binary semantic segmentation with a U-Net. A challenge is the extreme class imbalance within the image. Preliminary results indicate that the method is valid to detect curves, yet more work is needed to single out the direct path TDOA with confidence.

Tree canopy height is one of the most important indicators of forest biomass, productivity, and ecosystemstructure, but it is challenging to measure accurately from the ground and from space. Here, we used a U-Netmodel adapted for regression to map the canopy height of all trees in the state of California with very highresolution aerial imagery 0.6 m from the USDA-NAIP program. The U-Net model was trained using canopyheight models computed from aerial LiDAR data as a reference, along with corresponding RGB-NIR NAIP imagescollected in 2020. We evaluated the performance of the deep-learning model using 42 independent 1 km2 areasacross various forest types and landscape variations in California. Our predictions of tree heights exhibited amean error of 2.9 m and showed relatively low systematic bias across the entire range of tree heights present inCalifornia. In 2020, trees taller than 5 m covered ∼ 19.3% of California. Our model successfully estimatedcanopy heights up to 50 m without saturation, outperforming existing canopy height products from globalmodels. The approach we used allowed for the reconstruction of the three-dimensional structure of individualtrees as observed from nadir-looking optical airborne imagery, suggesting a relatively robust estimation andmapping capability, even in the presence of image distortion. These findings demonstrate the potential of largescale mapping and monitoring of tree height, as well as potential biomass estimation, using NAIP imagery., Tree canopy height is one of the most important indicators of forest biomass, productivity, and ecosystem structure, but it is challenging to measure accurately from the ground and from space. Here, we used a U-Net model adapted for regression to map the canopy height of all trees in the state of California with very high-resolution aerial imagery 0.6 m from the USDA-NAIP program. The U-Net model was trained using canopy height models computed from aerial LiDAR data as a reference, along with corresponding RGB-NIR NAIP images collected in 2020. We evaluated the performance of the deep-learning model using 42 independent 1 km2 areas across various forest types and landscape variations in California. Our predictions of tree heights exhibited a mean error of 2.9 m and showed relatively low systematic bias across the entire range of tree heights present in California. In 2020, trees taller than 5 m covered ∼ 19.3% of California. Our model successfully estimated canopy heights up to 50 m without saturation, outperforming existing canopy height products from global models. The approach we used allowed for the reconstruction of the three-dimensional structure of individual trees as observed from nadir-looking optical airborne imagery, suggesting a relatively robust estimation and mapping capability, even in the presence of image distortion. These findings demonstrate the potential of large-scale mapping and monitoring of tree height, as well as potential biomass estimation, using NAIP imagery.

In intensively managed forests in Europe, where forests are divided into stands of small size and may show heterogeneity within stands, a high spatial resolution (10–––20 m) is needed to capture the differences in canopy height. In this work, we developed a deep learning model based on multi-sensor remote sensing measurements to create a high-resolution canopy height map over the “Landes de Gascogne” forest in France, a large maritime pine plantation of 13,000 km2 with flat terrain and intensive management. This area is characterized by even-aged and mono-specific stands, of a typical length of a few hundred meters, harvested every 35 to 50 years. Our deep learning U-Net model uses multi-band images from Sentinel-1 and Sentinel-2 with composite time averages as input to predict tree height derived from GEDI waveforms. The evaluation is performed with external validation data from forest inventory plots and a stereo 3D reconstruction model based on Skysat imagery available at specific locations. We trained seven different U-Net models based on combinations of Sentinel-1 and Sentinel-2 bands to evaluate the importance of each sensor in the dominant height retrieval. The model outputs allow us to generate a 10 m resolution canopy height map of the whole “Landes de Gascogne” forest area for 2020 with a mean absolute error of 2.02 m on the test dataset. The best predictions were obtained using all available bands from Sentinel-1 and Sentinel-2 but using only one satellite source also provided good predictions. For all validation datasets in coniferous forests, our model showed better metrics than previous canopy height models available in the same region., In intensively managed forests in Europe, where forests are divided into stands of small size and may show heterogeneity within stands, a high spatial resolution (10–––20 m) is needed to capture the differences in canopy height. In this work, we developed a deep learning model based on multi-sensor remote sensing measurements to create a high-resolution canopy height map over the “Landes de Gascogne” forest in France, a large maritime pine plantation of 13,000 km2 with flat terrain and intensive management. This area is characterized by even-aged and mono-specific stands, of a typical length of a few hundred meters, harvested every 35 to 50 years. Our deep learning U-Net model uses multi-band images from Sentinel-1 and Sentinel-2 with composite time averages as input to predict tree height derived from GEDI waveforms. The evaluation is performed with external validation data from forest inventory plots and a stereo 3D reconstruction model based on Skysat imagery available at specific locations. We trained seven different U-Net models based on combinations of Sentinel-1 and Sentinel-2 bands to evaluate the importance of each sensor in the dominant height retrieval. The model outputs allow us to generate a 10 m resolution canopy height map of the whole “Landes de Gascogne” forest area for 2020 with a mean absolute error of 2.02 m on the test dataset. The best predictions were obtained using all available bands from Sentinel-1 and Sentinel-2 but using only one satellite source also provided good predictions. For all validation datasets in coniferous forests, our model showed better metrics than previous canopy height models available in the same region.

Tree canopy height is one of the most important indicators of forest biomass, productivity, and ecosystemstructure, but it is challenging to measure accurately from the ground and from space. Here, we used a U-Netmodel adapted for regression to map the canopy height of all trees in the state of California with very highresolution aerial imagery 0.6 m from the USDA-NAIP program. The U-Net model was trained using canopyheight models computed from aerial LiDAR data as a reference, along with corresponding RGB-NIR NAIP imagescollected in 2020. We evaluated the performance of the deep-learning model using 42 independent 1 km2 areasacross various forest types and landscape variations in California. Our predictions of tree heights exhibited amean error of 2.9 m and showed relatively low systematic bias across the entire range of tree heights present inCalifornia. In 2020, trees taller than 5 m covered ∼ 19.3% of California. Our model successfully estimatedcanopy heights up to 50 m without saturation, outperforming existing canopy height products from globalmodels. The approach we used allowed for the reconstruction of the three-dimensional structure of individualtrees as observed from nadir-looking optical airborne imagery, suggesting a relatively robust estimation andmapping capability, even in the presence of image distortion. These findings demonstrate the potential of largescale mapping and monitoring of tree height, as well as potential biomass estimation, using NAIP imagery., Tree canopy height is one of the most important indicators of forest biomass, productivity, and ecosystem structure, but it is challenging to measure accurately from the ground and from space. Here, we used a U-Net model adapted for regression to map the canopy height of all trees in the state of California with very high-resolution aerial imagery 0.6 m from the USDA-NAIP program. The U-Net model was trained using canopy height models computed from aerial LiDAR data as a reference, along with corresponding RGB-NIR NAIP images collected in 2020. We evaluated the performance of the deep-learning model using 42 independent 1 km2 areas across various forest types and landscape variations in California. Our predictions of tree heights exhibited a mean error of 2.9 m and showed relatively low systematic bias across the entire range of tree heights present in California. In 2020, trees taller than 5 m covered ∼ 19.3% of California. Our model successfully estimated canopy heights up to 50 m without saturation, outperforming existing canopy height products from global models. The approach we used allowed for the reconstruction of the three-dimensional structure of individual trees as observed from nadir-looking optical airborne imagery, suggesting a relatively robust estimation and mapping capability, even in the presence of image distortion. These findings demonstrate the potential of large-scale mapping and monitoring of tree height, as well as potential biomass estimation, using NAIP imagery.

In intensively managed forests in Europe, where forests are divided into stands of small size and may show heterogeneity within stands, a high spatial resolution (10–––20 m) is needed to capture the differences in canopy height. In this work, we developed a deep learning model based on multi-sensor remote sensing measurements to create a high-resolution canopy height map over the “Landes de Gascogne” forest in France, a large maritime pine plantation of 13,000 km2 with flat terrain and intensive management. This area is characterized by even-aged and mono-specific stands, of a typical length of a few hundred meters, harvested every 35 to 50 years. Our deep learning U-Net model uses multi-band images from Sentinel-1 and Sentinel-2 with composite time averages as input to predict tree height derived from GEDI waveforms. The evaluation is performed with external validation data from forest inventory plots and a stereo 3D reconstruction model based on Skysat imagery available at specific locations. We trained seven different U-Net models based on combinations of Sentinel-1 and Sentinel-2 bands to evaluate the importance of each sensor in the dominant height retrieval. The model outputs allow us to generate a 10 m resolution canopy height map of the whole “Landes de Gascogne” forest area for 2020 with a mean absolute error of 2.02 m on the test dataset. The best predictions were obtained using all available bands from Sentinel-1 and Sentinel-2 but using only one satellite source also provided good predictions. For all validation datasets in coniferous forests, our model showed better metrics than previous canopy height models available in the same region., In intensively managed forests in Europe, where forests are divided into stands of small size and may show heterogeneity within stands, a high spatial resolution (10–––20 m) is needed to capture the differences in canopy height. In this work, we developed a deep learning model based on multi-sensor remote sensing measurements to create a high-resolution canopy height map over the “Landes de Gascogne” forest in France, a large maritime pine plantation of 13,000 km2 with flat terrain and intensive management. This area is characterized by even-aged and mono-specific stands, of a typical length of a few hundred meters, harvested every 35 to 50 years. Our deep learning U-Net model uses multi-band images from Sentinel-1 and Sentinel-2 with composite time averages as input to predict tree height derived from GEDI waveforms. The evaluation is performed with external validation data from forest inventory plots and a stereo 3D reconstruction model based on Skysat imagery available at specific locations. We trained seven different U-Net models based on combinations of Sentinel-1 and Sentinel-2 bands to evaluate the importance of each sensor in the dominant height retrieval. The model outputs allow us to generate a 10 m resolution canopy height map of the whole “Landes de Gascogne” forest area for 2020 with a mean absolute error of 2.02 m on the test dataset. The best predictions were obtained using all available bands from Sentinel-1 and Sentinel-2 but using only one satellite source also provided good predictions. For all validation datasets in coniferous forests, our model showed better metrics than previous canopy height models available in the same region.

Aquest treball busca abordar un dels problemes més persistents en la gravació i reproducció d'àudio: la presència de distorsions no desitjades que degraden la qualitat i la fidelitat del so. Aquest treball de fi de grau es proposa desenvolupar i avaluar diferents tècniques per eliminar o reduir la distorsió d'àudio. Aquestes tècniques tenen com a objectiu millorar la qualitat del so al reduir o eliminar sorolls no desitjats, reverberacions, interferències i altres tipus de distorsions que afecten negativament l'experiència auditiva., This work seeks to address one of the most persistent problems in audio recording and playback: the presence of unwanted distortions that degrade sound quality and fidelity. This final thesis aims to develop and evaluate different audio enhancement techniques.These techniques aim to improve sound quality by reducing or eliminating unwanted noise, reverberation, interference and other types of distortions that negatively affect the listening experience.

Cloud is an aerosol consisting of visible mass of miniature liquid droplets, frozen crystal, or other particles suspended in the atmosphere. The study of atmospheric clouds is crucial for us to better understand and predict the behaviors of clouds, which has implications for climate, weather, aviation safety, agriculture, and energy production. Convolutional neural network (CNN) method is applied to train an atmospheric cloud image detection model to identify the presence of cloud and classify them. Supervised learning method is applied to train the model such that the machine is given labeled cloud image dataset to learn how to classify and predict the presence of cloud. U-Net architecture is used to train the atmospheric cloud image detection model because the architecture has the highest performance in image segmentation especially object detection in satellite images. The 38-Cloud Dataset which is used to train the model, is obtained from Landsat 8 Earth observation satellite. The dataset is randomly divided into training set (75% of the total images) and validation set (25% of the total images). Following this, the dataset is preprocessed and transformed into tensors to train the model. The training has been carried out for 50 epochs. Apart from the U-Net architecture proposed, the architecture is further modified with ResNet34 and VGG16 and the performance of each model is studied. The recognition accuracy obtained for atmospheric cloud image detection trained with the dataset achieved 97%. With this accuracy, U-Net architecture can be justified as a powerful and suitable convolutional neural network in performing atmospheric cloud image detection.&nbsp

The ability to quantify uncertainties associated with neural network predictions is crucial when they are relied upon in decision-making processes, especially in safety-critical applications like radiation therapy. In this paper, a single-model estimator of both epistemic and aleatoric uncertainties in a regression 3D U-net used for radiation dose prediction is presented. To capture epistemic uncertainty, Monte Carlo Dropout is employed, leveraging dropout during test-time inference to obtain a distribution of predictions. The variability among these predictions is used to estimate the model’s epistemic uncertainty. For quantifying aleatoric uncertainty quantile regression, which models conditional quantiles of the output distribution, is used. The method enables the estimation of prediction intervals of a user-specified significance level, where the difference between the upper and lower bound of the interval quantifies the aleatoric uncertainty. The proposed approach is evaluated on two datasets of prostate and breast cancer patient geometries and corresponding radiation doses. Results demonstrate that the quantile regression method provides well-calibrated prediction intervals, allowing for reliable aleatoric uncertainty estimation. Furthermore, the epistemic uncertainty obtained through Monte Carlo Dropout proves effective in identifying out-of-distribution examples, highlighting its usefulness for detecting anomalous cases where the model makes uncertain predictions., Förmågan att kvantifiera osäkerheter i samband med neurala nätverksprediktioner är avgörande när de åberopas i beslutsprocesser, särskilt i säkerhetskritiska tillämpningar såsom strålterapi. I denna rapport presenteras en en-modellsimplementation för att uppskatta både epistemiska och aleatoriska osäkerheter i ett 3D regressions-U-net som används för att prediktera stråldos. För att fånga epistemisk osäkerhet används Monte Carlo Dropout, som utnyttjar dropout under testtidsinferens för att få en fördelning av prediktioner. Variabiliteten mellan dessa prediktioner används för att uppskatta modellens epistemiska osäkerhet. För att kvantifiera den aleatoriska osäkerheten används kvantilregression, eller quantile regression, som modellerar de betingade kvantilerna i outputfördelningen. Metoden möjliggör uppskattning av prediktionsintervall med en användardefinierad signifikansnivå, där skillnaden mellan intervallets övre och undre gräns kvantifierar den aleatoriska osäkerheten. Den föreslagna metoden utvärderas på två dataset innehållandes geometrier för prostata- och bröstcancerpatienter och korresponderande stråldoser. Resultaten visar på att kvantilregression ger välkalibrerade prediktionsintervall, vilket tillåter en tillförlitlig uppskattning av den aleatoriska osäkerheten. Dessutom visar sig den epistemiska osäkerhet som erhålls genom Monte Carlo Dropout vara användbar för att identifiera datapunkter som inte tillhör samma fördelning som träningsdatan, vilket belyser dess lämplighet för att upptäcka avvikande datapunkter där modellen gör osäkra prediktioner.

The first Metatarsophalangeal (MTP) joint is essential for foot biomechanics and weight-bearing activities. Osteoarthritis in this joint can lead to pain, discomfort, and limited mobility. In order to treat this, Episurf Medical is working to produce individualized implants based on 3D segmentations of the joint. As manual segmentations are both time- and cost-consuming, and susceptible to human errors, automatic approaches are preferred. This thesis uses U-Net and DeepEdit as deep-learning based methods for segmentation of the MTP joint, with the latter being evaluated with and without user interactions. The dataset used in this study consisted of 38 CT images, where each model was trained on 30 images, and the remaining images were used as a test set. The final models were evaluated and compared with regards to the Dice Similarity Coefficient (DSC), precision, and recall. The U-Net model achieved DSC 0.944, precision 0.961, and recall 0.929. The automatic DeepEdit approach obtained DSC of 0.861, precision of 0.842, and recall of 0.891, while the interactive DeepEdit approach resulted in DSC of 0.918, precision of 0.912, and recall of 0.928. All pairwise comparisons in terms of precision and DSC showed significant differences (p<0.05), where U-Net had the highest performance, while the difference in recall was not found to be significant (p>0.05) for any comparison. The lower performances of DeepEdit compared to U-Net could be due to lower spatial resolution in the segmentations. Nevertheless, DeepEdit remains a promising method, and further investigations of unexplored areas could be addressed as future work., Den första Metatarsalphalangeal(MTP) leden är viktig för fotens biomekanik och viktbärande aktiviteter. Artros i denna led kan leda till smärta, obehag och begränsad rörlighet. För att behandla detta arbetar Episurf Medical med att producera individanpassade implantat baserat på 3D segmenteringar av leden. Då manuella segmenteringar både är tids- och kostnadskrävande, samt känsliga för mänskliga fel, föredras automatiska metoder. Denna avhandling använder U-Net och DeepEdit som djupinlärningsbaserade metoder för segm- entering av MTP leden, varav det senare utvärderas med och utan användarint- eraktion. Datasetet som användes i denna studie bestod av 38 CT bilder, där varje modell tränades på 30 bilder och de återstående användes som testdata. De slutliga modellerna utvärderades och jämfördes med avseende på Dice Similarity Coefficient (DSC), precision och recall. U-Net modellen uppnådde DSC 0.944, precision 0.961 och recall 0.929. Den automatiska DeepEdit metoden erhöll DSC 0.861, precision 0.842 och recall 0.891, medan den interaktiva DeepEdit metoden resulterade i DSC 0.918, precision 0.912 och recall 0.928. Alla parvisa jämförelser avseende precision och DSC visade signifikanta skillnader (p<0.05), där U-Net hade den högsta prestandan, medan skillnaden i recall inte visade sig vara signifikant (p>0.05) för någon jämförelse. Den lägre prestandan för DeepEdit jämfört med U-Net kan bero på lägre spatiell upplösning i segmenteringarna. Dock är DeepEdit fortfarande en lovande metod, och ytterligare undersökningar av outforskade områden kan tas upp som framtida arbete.

© ACM 2023. This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in Advances in Soft Computing: 22nd Mexican International Conference on Artificial Intelligence, MICAI 2023, Yucatán, Mexico, November 13-18, 2023, proceedings, part IIhttp://dx.doi.org/10.1007/978-3-031-47640-2_14, This contribution presents a deep learning method for the segmentation of prostate zones in MRI images based on U-Net using additive and feature pyramid attention modules, which can improve the workflow of prostate cancer detection and diagnosis. The proposed model is compared to seven different U-Net-based architectures. The automatic segmentation performance of each model of the central zone (CZ), peripheral zone (PZ), transition zone (TZ) and Tumor were evaluated using Dice Score (DSC), and the Intersection over Union (IoU) metrics. The proposed alternative achieved a mean DSC of 84.15% and IoU of 76.9% in the test set, outperforming most of the studied models in this work except from R2U-Net and attention R2U-Net architectures., The authors wish to acknowledge the Mexican Council for Science and Technology (CONACYT) for the support in terms of post graduate scholarships in this project, and the Data Science Hub at Tecnologico de Monterrey for their support on this project., Peer Reviewed, Postprint (author's final draft)

Intravenous (IV) access is an important daily clinical procedure that delivers fluids or medication into a patient’s vein. However, IV insertion is very challenging where clinicians are suffering in locating the subcutaneous vein due to patients’ physiological factors such as hairy forearm and thick dermis fat, and also medical staff’s level of fatigue. To resolve this issue, researchers have proposed autonomous machines to be used for IV access, but such equipment are lacking capability in detecting the vein accurately. Therefore, this project proposes an automatic vein detection algorithm using deep learning for IV access purpose. U-Net, a fully connected network (FCN) architecture is employed in this project due to its capability in detecting the near-infrared (NIR) subcutaneous vein. Data augmentation is applied to increase the dataset size and reduce the bias from overfitting. The original U-Net architecture is optimized by replacing up-sampling with transpose convolution as well as the additional implementation of batch normalization besides reducing the number of layers to diminish the risk of overfitting. After fine-tuning and retraining the hypermodel, an unsupervised dataset is used to evaluate the hypermodel by selecting 10 checkpoints for each forearm image and comparing the checkpoints on predicted outputs to determine true positive vein pixels. The proposed lightweight U-Net has achieved slightly lower accuracy (0.8871) than the original U-Net architecture. Even so, the sensitivity, specificity, and precision are greatly improved by achieving 0.7806, 0.9935, and 0.9918 respectively. This result indicates that the proposed algorithm can be applied into the venipuncture machine to accurately locate the subcutaneous vein for intravenous (IV) procedures.

Background Manual analysis of (mini-)rhizotron (MR) images is tedious. Several methods have been proposed for semantic root segmentation based on homogeneous, single-source MR datasets. Recent advances in deep learning (DL) have enabled automated feature extraction, but comparisons of segmentation accuracy, false positives and transferability are virtually lacking. Here we compare six state-of-the-art methods and propose two improved DL models for semantic root segmentation using a large MR dataset with and without augmented data. We determine the performance of the methods on a homogeneous maize dataset, and a mixed dataset of > 8 species (mixtures), 6 soil types and 4 imaging systems. The generalisation potential of the derived DL models is determined on a distinct, unseen dataset. Results The best performance was achieved by the U-Net models; the more complex the encoder the better the accuracy and generalisation of the model. The heterogeneous mixed MR dataset was a particularly challenging for the non-U-Net techniques. Data augmentation enhanced model performance. We demonstrated the improved performance of deep meta-architectures and feature extractors, and a reduction in the number of false positives. Conclusions Although correction factors are still required to match human labelled root lengths, neural network architectures greatly reduce the time required to compute the root length. The more complex architectures illustrate how future improvements in root segmentation within MR images can be achieved, particularly reaching higher segmentation accuracies and model generalisation when analysing real-world datasets with artefacts-limiting the need for model retraining.

Background Manual analysis of (mini-)rhizotron (MR) images is tedious. Several methods have been proposed for semantic root segmentation based on homogeneous, single-source MR datasets. Recent advances in deep learning (DL) have enabled automated feature extraction, but comparisons of segmentation accuracy, false positives and transferability are virtually lacking. Here we compare six state-of-the-art methods and propose two improved DL models for semantic root segmentation using a large MR dataset with and without augmented data. We determine the performance of the methods on a homogeneous maize dataset, and a mixed dataset of > 8 species (mixtures), 6 soil types and 4 imaging systems. The generalisation potential of the derived DL models is determined on a distinct, unseen dataset. Results The best performance was achieved by the U-Net models; the more complex the encoder the better the accuracy and generalisation of the model. The heterogeneous mixed MR dataset was a particularly challenging for the non-U-Net techniques. Data augmentation enhanced model performance. We demonstrated the improved performance of deep meta-architectures and feature extractors, and a reduction in the number of false positives. Conclusions Although correction factors are still required to match human labelled root lengths, neural network architectures greatly reduce the time required to compute the root length. The more complex architectures illustrate how future improvements in root segmentation within MR images can be achieved, particularly reaching higher segmentation accuracies and model generalisation when analysing real-world datasets with artefacts-limiting the need for model retraining., info:eu-repo/semantics/openAccess

Alzheimer’s Disease is a progressive and fatal neurodegenerative disease which affects millions of people around the world. The pathophysiology of the disease is characterized by the accumulation of neuritic amyloid plaques and neurofibrillary tau tangles within the hippocampus and many surrounding structures. Tau tangles, in particular, are commonly used to identify the stage of disease progression. Currently, only the presence or absence of tau tangles, together with simple staging information, in specific brain regions is noted at autopsy. An improvement to this approach is to measure the distribution of tau tangles across large brain samples in order to better characterize the progression of the disease. In this work, we build a framework for comparing the ability of different machine learning models to identify the locations of tau tangles in postmortem neural microscopy images. In particular, we focus on the development of a set of software tools which transform probability heatmaps into a set of region proposals for all the tau tangles within an image. We then construct two different machine learning models and compare their performance using a precision-recall (PR) framework.

Minor release for Zenodo archiving. Code for DRCNN: Hyperspectral data reduction in supervised deep learning with Data Reduction Convolutional Neural Network For a full description, please visit https://github.com/mzeegers/DRCNN

Minor release for Zenodo archiving. Code for DeepFODDataGenerator: A workflow for generating deep learning data for X-ray based object detection For a full description, please visit https://github.com/mzeegers/DeepFODDataGenerator

Magnetic resonance imaging (MRI) is a commonly used non-invasive imaging technique that provides high resolution images of soft tissue. One problem with MRI is that it is sensitive to signal quality issues. The issues can arise for various reasons, for example by metal located either inside or outside of the body. Another common signal quality issue is caused by the patient being partly placed outside field of view of the MRI scanner. This thesis aims to investigate the possibility to automatically detect these signal quality issues using deep neural networks. More specifically, two different 3D CNN network types were studied, a classification-based approach and a reconstruction-based approach. The datasets used consist of MRI volumes from UK Biobank which have been processed and manually annotated by operators at AMRA Medical. For the classification method four different network architectures were explored utilising supervised learning with multi-label classification. The classification method was evaluated using accuracy and label-based evaluation metrics, such as macro-precision, macro-recall and macro-F1. The reconstruction method was based on anomaly detection using an autoencoder which was trained to reconstruct volumes without any artefacts. A mean squared prediction error was calculated for the reconstructed volume and compared against a threshold in order to classify a volume with or without artefacts. The idea was that volumes containing artefacts should be more difficult to reconstruct and thus, result in a higher prediction error. The reconstruction method was evaluated using accuracy, precision, recall and F1-score. The results show that the classification method has overall higher performance than the reconstruction method. The achieved accuracy for the classification method was 98.0% for metal artefacts and 97.5% for outside field of view artefacts. The best architecture for the classification method proved to be DenseNet201. The reconstruction met

For manufacturers in the wood industry, an important way to make the production more effective is to automate the process of detecting defects and different attributes on boards. One important attribute on most boards is heartwood and sapwood. This thesis project was conducted at the company MiCROTEC and aims to investigate methods to classify heartwood and sapwood on boards. The dataset used in this project consisted of oak boards. In order to increase the amount of information retrieved from the boards, hyperspectral imaging was used instead of conventional RGB cameras. Based on this data, deep learning models in the form of U-Net and U-within-U-Net architecture as well as different spectral dimensionality reduction methods were developed to segment boards in heartwood and sapwood. The performance of these deep learning models was compared to PLS-DA and SVM. PLS-DA has already been used at MiCROTEC and has been used in this work for comparison as a baseline model. The result of the thesis work showed that a deep learning approach could increase the F1-Score from 0.730 for the baseline classifier PLS-DA to an F1-Score of 0.918, and that the different spectral reduction methods only had a small impact on the result. The increase in F1-score was mainly due to an increase in precision, since the PLS-DA had a similar recall as the deep learning models.

Magnetic resonance imaging (MRI) is a commonly used non-invasive imaging technique that provides high resolution images of soft tissue. One problem with MRI is that it is sensitive to signal quality issues. The issues can arise for various reasons, for example by metal located either inside or outside of the body. Another common signal quality issue is caused by the patient being partly placed outside field of view of the MRI scanner. This thesis aims to investigate the possibility to automatically detect these signal quality issues using deep neural networks. More specifically, two different 3D CNN network types were studied, a classification-based approach and a reconstruction-based approach. The datasets used consist of MRI volumes from UK Biobank which have been processed and manually annotated by operators at AMRA Medical. For the classification method four different network architectures were explored utilising supervised learning with multi-label classification. The classification method was evaluated using accuracy and label-based evaluation metrics, such as macro-precision, macro-recall and macro-F1. The reconstruction method was based on anomaly detection using an autoencoder which was trained to reconstruct volumes without any artefacts. A mean squared prediction error was calculated for the reconstructed volume and compared against a threshold in order to classify a volume with or without artefacts. The idea was that volumes containing artefacts should be more difficult to reconstruct and thus, result in a higher prediction error. The reconstruction method was evaluated using accuracy, precision, recall and F1-score. The results show that the classification method has overall higher performance than the reconstruction method. The achieved accuracy for the classification method was 98.0% for metal artefacts and 97.5% for outside field of view artefacts. The best architecture for the classification method proved to be DenseNet201. The reconstruction met

A brain tumor is an abnormal growth of cells within the brain, which can be categorized into primary and secondary tumor types. The most common type of primary tumors in adults are gliomas, which can be further classified into high-grade gliomas (HGGs) and low-grade gliomas (LGGs). Approximately 50% of patients diagnosed with HGG pass away within 1-2 years. Therefore, the early detection and prompt treatment of brain tumors are essential for effective management and improved patient outcomes. Brain tumor segmentation is a task in medical image analysis that entails distinguishing brain tumors from normal brain tissue in magnetic resonance imaging (MRI) scans. Computer vision algorithms and deep learning models capable of analyzing medical images can be leveraged for brain tumor segmentation. These algorithms and models have the potential to provide automated, reliable, and non-invasive screening for brain tumors, thereby enabling earlier and more effective treatment. For a considerable time, Convolutional Neural Networks (CNNs), including the U-Net, have served as the standard backbone architectures employed to address challenges in computer vision. In recent years, the Transformer architecture, which already has firmly established itself as the new state-of-the-art in the field of natural language processing (NLP), has been adapted to computer vision tasks. The Vision Transformer (ViT) and the Swin Transformer are two architectures derived from the original Transformer architecture that have been successfully employed for image analysis. The emergence of Transformer based architectures in the field of computer vision calls for an investigation whether CNNs can be rivaled as the de facto architecture in this field. This thesis compares the performance of four model architectures, namely the Swin Transformer, the Vision Transformer, the 2D U-Net, and the 2D U-Net which is implemented with the nnU-Net framework. These model architectures are trained using increasing

Purpose: The diagnosis and prognosis of breast cancer relies on histopathology image analysis. In this context, proliferation markers, especially Ki67, are increasingly important. The diagnosis using these markers is based on the quantification of proliferation, which implies the counting of Ki67 positive and negative tumoral cells in epithelial regions, thus excluding stromal cells. However, stromal cells are often very difficult to distinguish from negative tumoral cells in Ki67 images and often lead to errors when automatic analysis is used. Approach: We study the use of automatic semantic segmentation based on convolutional neural networks (CNNs) to separate stromal and epithelial areas on Ki67 stained images. CNNs need to be accurately trained with extensive databases with associated ground truth. As such databases are not publicly available, we propose a method to produce them with minimal manual labelling effort. Inspired by the procedure used by pathologists, we have produced the database relying on knowledge transfer from cytokeratin-19 images to Ki67 using an image-to-image (I2I) translation network. Results: The automatically produced stroma masks are manually corrected and used to train a CNN that predicts very accurate stroma masks for unseen Ki67 images. An F-score value of 0.87 is achieved. Examples of effect on the KI67 score show the importance of the stroma segmentation. Conclusions: An I2I translation method has proved very useful for building ground-truth labeling in a task where manual labeling is unfeasible. With reduced correction effort, a dataset can be built to train neural networks for the difficult problem of separating epithelial regions from stroma in stained images where separation is very hard without additional information., This work was supported by the Spanish Research Agency (AEI) (Project No. PID2020-116907RB-I00) of the call MCIN/AEI/10.13039/50110001103 and the European Regional Development Funds under program FEDER Catalunya 2014–2020 and SA18-014623 DIGIPATICS., Peer Reviewed, Postprint (published version)

In this thesis, application of imaging in assessment of zebrafish (zf) cardiology is discussed. Medical imaging seeks to expose internal structures, as well as to diagnose and treat disease. Medical imaging also establishes a database of normal anatomy and physiology to make it possible to identify abnormalities. Considering that the embryonic zebrafish is transparent, bright field microscopic videos could reveal heart mechanism and could be useful for quantification of it, although microscopic imaging can be useful for adult zebrafish as well. Different imaging methods used in other works is discussed first. Later, the cardiovascular parameters that can be measured using imaging are defined. We then compare different digital image processing and deep learning algorithms that have been employed to process or segment images from zebrafish. At the end of the chapter challenges mutant type are investigated.

Previous studies have shown that Synthetic Aperture Radar (SAR) is able to detect burned areas, serving as a key data source for monitoring active wildfires in situations where optical sensors are hindered by dense smoke or cloud cover. Radar remote sensing provides rich and useful data on historical burned areas, which are critical for large-scale wildfire burned area mapping. This study aims to unveil the potential inherent in Sentinel-1 C-Band SAR data and to investigate the impact of reference masks on wildfire burned area mapping. SAR images frequently exhibit disruptive speckle noise, which manifests as isolated pixels or small regions, thereby hindering the seamless identification of burned areas through SAR data. To mitigate noise and enhance the connectivity of SAR-based burned area delineation, we propose a novel approach: a Total-Variation (TV) regularized U-Net model, tailored to learn from noisy pseudo masks derived from Sentinel-1 C-Band SAR backscattering data. To validate its efficacy, we assembled a dataset encompassing 16 geographically diverse wildfire events, incorporating SAR-based kMap, a SAR-based binary label (designated as SARREF), and an optical-based binary label (referred to as OptREF). The kMap, serving as a relative indicator of SAR backscattering change, is defined as the divergence between post-fire SAR backscattering and the temporal average, normalized by the temporal standard deviation of pre-fire SAR time series. Utilizing SAR-based kMap as input and SARREF or OptREF as labels, we systematically explored the impact of TV regularization on SAR-based burned area mapping. Our experimental findings highlight several key insights: 1) Vanilla U-Net considerably elevates the F1 score from 0.61 to 0.66 when compared to threshold-based approaches; 2) The TV-regularized U-Net exhibits a notable capacity for enhancing both accuracy and the connectivity of SAR-based burned areas. When trained with noisy SARREF, it achieves a substantial F1 sc, QC 20230830

Current disaster response strategies are based on damage assessments carried out on the ground, which can be dangerous following a ä destructive event. Damage assessments can also be performed remotely using satellite imagery, but are usually carried out through visual interpretation, which can take a lot of time. This thesis explored a way of using artificial intelligence to automate remote damage assessment. We implemented a dual-task U-Net deep learning model, trained it with the xBD dataset for assessing building damage, and applied the model to pre- and post-event very high resolution satellite imagery of the February 6, 2023 earthquakes in Turkey. The results were compared to damage maps produced using a traditional object based method by calculating the F1 scores associated with the outputs of each method and ground truth data that we compiled. The study areas were parts of the two cities Kahramanmaraş and Antakya. The deep learning model almost only correctly identified undamaged buildings, achieving F1 scores of 0.95 during training as well as 0.93 and 0.83 in the damage assessments of Kahramanmaras and Antakya, respectively. For the other damage classes, the best result was the classification of destroyed buildings, both in training and in the study areas, with a F1-score of 0.45 in training and 0.16 in Kahramanmaraş. The deep learning model performed similarly to the object based method. Although the thesis did not yield good damage maps in the areas of interest, it had many limitations, and there is still a lot of potential for deep learning models to be useful in building damage assessment.

Remote Sensing (RS) technology provides valuable information about Earth by enabling an overview of the planet from above, making it a much-needed resource for many applications. Given the abundance of RS data and continued urbanisation, there is a need for efficient approaches to leverage RS data and its unique characteristics for the assessment and management of urban areas. Consequently, employing Deep Learning (DL) for RS applications has attracted much attention over the past few years. In this thesis, novel datasets consisting of satellite RS images over urban areas in Sweden were compiled from Sentinel-2 multispectral, Sentinel-1 Synthetic Aperture Radar (SAR) and Urban Atlas 2018 Land Use/Land Cover (LULC) data. Then, DL was applied for multiband and multiclass semantic image segmentation of LULC. The contributions of complementary spectral, temporal and SAR data and spectral indices to LULC classification performance compared to using only Sentinel-2 data with red, green and blue spectral bands were investigated by implementing DL models based on the fully convolutional network-based architecture, U-Net, and performing data fusion. Promising results were achieved with 25 possible LULC classes. Furthermore, almost all DL models at an overall model level and all DL models at an individual class level for most LULC classes benefited from complementary satellite RS data with varying degrees of classification improvement. Additionally, practical knowledge and insights were gained from evaluating the results and are presented regarding satellite RS data characteristics and semantic segmentation of LULC in urban areas. The obtained results are helpful for practitioners and researchers applying or intending to apply DL for semantic segmentation of LULC in general and specifically in Swedish urban environments.

Remote Sensing (RS) technology provides valuable information about Earth by enabling an overview of the planet from above, making it a much-needed resource for many applications. Given the abundance of RS data and continued urbanisation, there is a need for efficient approaches to leverage RS data and its unique characteristics for the assessment and management of urban areas. Consequently, employing Deep Learning (DL) for RS applications has attracted much attention over the past few years. In this thesis, novel datasets consisting of satellite RS images over urban areas in Sweden were compiled from Sentinel-2 multispectral, Sentinel-1 Synthetic Aperture Radar (SAR) and Urban Atlas 2018 Land Use/Land Cover (LULC) data. Then, DL was applied for multiband and multiclass semantic image segmentation of LULC. The contributions of complementary spectral, temporal and SAR data and spectral indices to LULC classification performance compared to using only Sentinel-2 data with red, green and blue spectral bands were investigated by implementing DL models based on the fully convolutional network-based architecture, U-Net, and performing data fusion. Promising results were achieved with 25 possible LULC classes. Furthermore, almost all DL models at an overall model level and all DL models at an individual class level for most LULC classes benefited from complementary satellite RS data with varying degrees of classification improvement. Additionally, practical knowledge and insights were gained from evaluating the results and are presented regarding satellite RS data characteristics and semantic segmentation of LULC in urban areas. The obtained results are helpful for practitioners and researchers applying or intending to apply DL for semantic segmentation of LULC in general and specifically in Swedish urban environments.

Current disaster response strategies are based on damage assessments carried out on the ground, which can be dangerous following a ä destructive event. Damage assessments can also be performed remotely using satellite imagery, but are usually carried out through visual interpretation, which can take a lot of time. This thesis explored a way of using artificial intelligence to automate remote damage assessment. We implemented a dual-task U-Net deep learning model, trained it with the xBD dataset for assessing building damage, and applied the model to pre- and post-event very high resolution satellite imagery of the February 6, 2023 earthquakes in Turkey. The results were compared to damage maps produced using a traditional object based method by calculating the F1 scores associated with the outputs of each method and ground truth data that we compiled. The study areas were parts of the two cities Kahramanmaraş and Antakya. The deep learning model almost only correctly identified undamaged buildings, achieving F1 scores of 0.95 during training as well as 0.93 and 0.83 in the damage assessments of Kahramanmaras and Antakya, respectively. For the other damage classes, the best result was the classification of destroyed buildings, both in training and in the study areas, with a F1-score of 0.45 in training and 0.16 in Kahramanmaraş. The deep learning model performed similarly to the object based method. Although the thesis did not yield good damage maps in the areas of interest, it had many limitations, and there is still a lot of potential for deep learning models to be useful in building damage assessment.

The first Metatarsophalangeal (MTP) joint is essential for foot biomechanics and weight-bearing activities. Osteoarthritis in this joint can lead to pain, discomfort, and limited mobility. In order to treat this, Episurf Medical is working to produce individualized implants based on 3D segmentations of the joint. As manual segmentations are both time- and cost-consuming, and susceptible to human errors, automatic approaches are preferred. This thesis uses U-Net and DeepEdit as deep-learning based methods for segmentation of the MTP joint, with the latter being evaluated with and without user interactions. The dataset used in this study consisted of 38 CT images, where each model was trained on 30 images, and the remaining images were used as a test set. The final models were evaluated and compared with regards to the Dice Similarity Coefficient (DSC), precision, and recall. The U-Net model achieved DSC 0.944, precision 0.961, and recall 0.929. The automatic DeepEdit approach obtained DSC of 0.861, precision of 0.842, and recall of 0.891, while the interactive DeepEdit approach resulted in DSC of 0.918, precision of 0.912, and recall of 0.928. All pairwise comparisons in terms of precision and DSC showed significant differences (p<0.05), where U-Net had the highest performance, while the difference in recall was not found to be significant (p>0.05) for any comparison. The lower performances of DeepEdit compared to U-Net could be due to lower spatial resolution in the segmentations. Nevertheless, DeepEdit remains a promising method, and further investigations of unexplored areas could be addressed as future work., Den första Metatarsalphalangeal(MTP) leden är viktig för fotens biomekanik och viktbärande aktiviteter. Artros i denna led kan leda till smärta, obehag och begränsad rörlighet. För att behandla detta arbetar Episurf Medical med att producera individanpassade implantat baserat på 3D segmenteringar av leden. Då manuella segmenteringar både är tids- och kostnadskrävande, samt känsliga för mänskliga fel, föredras automatiska metoder. Denna avhandling använder U-Net och DeepEdit som djupinlärningsbaserade metoder för segm- entering av MTP leden, varav det senare utvärderas med och utan användarint- eraktion. Datasetet som användes i denna studie bestod av 38 CT bilder, där varje modell tränades på 30 bilder och de återstående användes som testdata. De slutliga modellerna utvärderades och jämfördes med avseende på Dice Similarity Coefficient (DSC), precision och recall. U-Net modellen uppnådde DSC 0.944, precision 0.961 och recall 0.929. Den automatiska DeepEdit metoden erhöll DSC 0.861, precision 0.842 och recall 0.891, medan den interaktiva DeepEdit metoden resulterade i DSC 0.918, precision 0.912 och recall 0.928. Alla parvisa jämförelser avseende precision och DSC visade signifikanta skillnader (p<0.05), där U-Net hade den högsta prestandan, medan skillnaden i recall inte visade sig vara signifikant (p>0.05) för någon jämförelse. Den lägre prestandan för DeepEdit jämfört med U-Net kan bero på lägre spatiell upplösning i segmenteringarna. Dock är DeepEdit fortfarande en lovande metod, och ytterligare undersökningar av outforskade områden kan tas upp som framtida arbete.

In the forensic studies of painting masterpieces, the analysis of the support is of major importance. For plain weave fabrics, the densities of vertical and horizontal threads are used as main features, while angle deviations from the vertical and horizontal axis are also of help. These features can be studied locally through the canvas. In this work, deep learning is proposed as a tool to perform these local densities and angle studies. We trained the model with samples from 36 paintings by Velázquez, Rubens or Ribera, among others. The data preparation and augmentation are dealt with at a first stage of the pipeline. We then focus on the supervised segmentation of crossing points between threads. The U-Net with inception and Dice loss are presented as good choices for this task. Densities and angles are then estimated based on the segmented crossing points. We report test results of the analysis of a few canvases and a comparison with methods in the frequency domain, widely used in this problem. We concluded that this new approach successes in some cases where the frequency analysis tools fail, while improves the results in others. Besides, our proposal does not need the labeling of part of the to be processed image. As case studies, we apply this novel algorithm to the analysis of two pairs of canvases by Velázquez and Murillo, to conclude that the fabrics used came from the same roll.

The ability to quantify uncertainties associated with neural network predictions is crucial when they are relied upon in decision-making processes, especially in safety-critical applications like radiation therapy. In this paper, a single-model estimator of both epistemic and aleatoric uncertainties in a regression 3D U-net used for radiation dose prediction is presented. To capture epistemic uncertainty, Monte Carlo Dropout is employed, leveraging dropout during test-time inference to obtain a distribution of predictions. The variability among these predictions is used to estimate the model’s epistemic uncertainty. For quantifying aleatoric uncertainty quantile regression, which models conditional quantiles of the output distribution, is used. The method enables the estimation of prediction intervals of a user-specified significance level, where the difference between the upper and lower bound of the interval quantifies the aleatoric uncertainty. The proposed approach is evaluated on two datasets of prostate and breast cancer patient geometries and corresponding radiation doses. Results demonstrate that the quantile regression method provides well-calibrated prediction intervals, allowing for reliable aleatoric uncertainty estimation. Furthermore, the epistemic uncertainty obtained through Monte Carlo Dropout proves effective in identifying out-of-distribution examples, highlighting its usefulness for detecting anomalous cases where the model makes uncertain predictions., Förmågan att kvantifiera osäkerheter i samband med neurala nätverksprediktioner är avgörande när de åberopas i beslutsprocesser, särskilt i säkerhetskritiska tillämpningar såsom strålterapi. I denna rapport presenteras en en-modellsimplementation för att uppskatta både epistemiska och aleatoriska osäkerheter i ett 3D regressions-U-net som används för att prediktera stråldos. För att fånga epistemisk osäkerhet används Monte Carlo Dropout, som utnyttjar dropout under testtidsinferens för att få en fördelning av prediktioner. Variabiliteten mellan dessa prediktioner används för att uppskatta modellens epistemiska osäkerhet. För att kvantifiera den aleatoriska osäkerheten används kvantilregression, eller quantile regression, som modellerar de betingade kvantilerna i outputfördelningen. Metoden möjliggör uppskattning av prediktionsintervall med en användardefinierad signifikansnivå, där skillnaden mellan intervallets övre och undre gräns kvantifierar den aleatoriska osäkerheten. Den föreslagna metoden utvärderas på två dataset innehållandes geometrier för prostata- och bröstcancerpatienter och korresponderande stråldoser. Resultaten visar på att kvantilregression ger välkalibrerade prediktionsintervall, vilket tillåter en tillförlitlig uppskattning av den aleatoriska osäkerheten. Dessutom visar sig den epistemiska osäkerhet som erhålls genom Monte Carlo Dropout vara användbar för att identifiera datapunkter som inte tillhör samma fördelning som träningsdatan, vilket belyser dess lämplighet för att upptäcka avvikande datapunkter där modellen gör osäkra prediktioner.

While the influence of surface melt on Antarctic ice shelf stability can be large, the duration and affected area of melt events are often small. Therefore, melt events are difficult to capture with remote sensing, as satellite sensors always face the trade-off between spatial and temporal resolution. To overcome this limitation, we developed UMelt: a surface melt record for all Antarctic ice shelves with a high spatial (500 m) and high temporal (12 h) resolution for the period 2016–2021. Our approach is based on a deep learning model, specifically a U-Net, which was developed in Google Earth Engine. The U-Net combines microwave remote sensing observations from three sources: Sentinel-1, Special Sensor Microwave Imager/Sounder (SSMIS), and Advanced Scatterometer (ASCAT). The U-Net was trained on the Shackleton Ice Shelf for melt seasons 2017–2021, using the fine-scale melt patterns of Sentinel-1 as reference data and SSMIS, ASCAT, a digital elevation model, and multi-year Sentinel-1 melt fraction as predictors. The trained U-Net performed well on the Shackelton Ice Shelf for test melt season 2016–2017 (accuracy: 91.3%; F1-score: 86.9%), and the Larsen C Ice Shelf, which was not considered during training (accuracy: 91.0%; F1-score: 89.3%). Using the trained U-Net model, we have successfully developed the UMelt record. UMelt allows Antarctic-wide surface melt to be detected at a small scale while preserving a high temporal resolution, which could lead to new insights into the response of ice shelves to a changing atmospheric forcing., Physical and Space Geodesy, Mathematical Geodesy and Positioning

This thesis introduces a Learned-Based Multi-View Semantic Stereo method, addressing the limitations of traditional and learned-based Multi-View Stereo (MVS) techniques in reconstructing reflective and low-textured regions, particularly prevalent in 3D models of buildings. Traditional methods lack completeness, while learned-based methods struggle with accuracy. Focusing on enhancing 3D models of buildings, this research integrates semantic information into the existing deep learning architecture for depth prediction, specifically CasMVSNet, to guide the reconstruction process. Three key strategies are employed: first, the incorporation of semantic maps into the network through a multi-modal approach; second, the introduction of a multi-modal refinement module at the end of the CasMVSNet model to improve the initial output depth maps; and third, the introduction of two new loss terms designed to enforce varying degrees of smoothness on specific semantic categories. Experimental results, conducted on the DTU dataset, demonstrate a significant enhancement in accuracy at the point cloud level while maintaining the completeness of the reconstructed models. Validation and generalization on the ETH3D dataset show consistent patterns. This research showcases the potential of integrating semantic guidance in 3D reconstruction of buildings, advancing the field of computer vision., Geomatics

Magnetic resonance electrical properties tomography is a type of quantitative magnetic resonance imaging that aims to reconstruct the conductivity and permittivity of biological tissue. These electrical properties of the tissue can be used to compute the specific absorption rate, to differentiate tumours from healthy tissue and for hyperthermia treatment planning. Several methods to reconstruct these electrical properties exist with different degrees of success. Combining analytical reconstruction methods with deep learning methods is left relatively unexplored in the field of magnetic resonance electrical properties tomography. Hence, this work explores such hybrid methods in which deep learning is embedded in an analytical reconstruction method. A recurrent inference machine is integrated into the iterative reconstruction scheme called Contrast Source Inversion, in an attempt to decrease its high computational load. Additionally, a U-net is trained to correct reconstructed conductivity maps using discrepancies in measured- and reconstructed phase data, which is based on the relation between conductivity and phase in the Helmholtz equation. The recurrent inference machine embedded version of contrast source inversion failed to achieve a desirable reconstruction quality with its current implementation. However, the large amount of potential improvements to its implementation motivates further research into its application before discarding it. The conductivity correction U-net is able to correct conductivity errors as small as 0.13 S/m when used iteratively or 0.05 S/m when used a single time when noiseless data is used. Further research in its capabilities of handling noisy data is required to assess practical usage., Electrical Engineering | Circuits and Systems

This thesis aims to automatically and reliably detect large-scale structures in turbulent flows. To achieve this, a U-net (a type of neural network) is trained using artificially generated data. From the network output, continuous structures are derived and general statistics, including, volume fraction, relative kinetic energy and length scales (using PCA) are computed. Detections were done on two homogenous isotropic turbulence (HIT) direct numerical simulation (DNS) datasets with Taylor Reynolds numbers of 175 and 1131, respectively. For both cases, the detected structures contained most of the volume and kinetic energy in the domain and were of the integral length scale. However, for the high Reynolds number, there were relatively half as many structures and the structures were roughly 4 times larger compared to those found in the low Reynolds number case., Aerospace Engineering