Search

Computational dentistry uses computerized methods and mathematical models for dental image analysis. One of the fundamental problems in computational dentistry is accurate tooth instance segmentation in high-resolution mesh data of intra-oral scans (IOS). This paper presents a new computational model based on deep neural networks, called Mask-MCNet, for end-to-end learning of tooth instance segmentation in 3D point cloud data of IOS. The proposed Mask-MCNet localizes each tooth instance by predicting its 3D bounding box and simultaneously segments the points that belong to each individual tooth instance. The proposed model processes the input raw 3D point cloud in its original spatial resolution without employing a voxelization or down-sampling technique. Such a characteristic preserves the finely detailed context in data like fine curvatures in the border between adjacent teeth and leads to a highly accurate segmentation as required for clinical practice (e.g. orthodontic planning). The experiments show that the Mask-MCNet outperforms state-of-the-art models by achieving 98% Intersection over Union (IoU) score on tooth instance segmentation which is very close to human expert performance.

Because the current Barrett's esophagus (BE) surveillance protocol suffers from sampling error of random biopsies and a high miss-rate of early neoplastic lesions, many new endoscopic imaging and sampling techniques have been developed. None of these techniques, however, have significantly increased the diagnostic yield of BE neoplasia. In fact, these techniques have led to an increase in the amount of visible information, yet endoscopists and pathologists inevitably suffer from variations in intra- and interobserver agreement. Artificial intelligence systems have the potential to overcome these endoscopist-dependent limitations.

Person re-identification (re-ID) aims to recognize an identity in non-overlapping camera views. Recently, re-ID received increased attention due to the growth of deep learning and its prominent applications in the field of automated video surveillance. The performance of deep learning-based methods relies heavily on the quality of training datasets and protocols. Particularly, parameters associated to the train and test set construction affect the overall performance. However, public re-ID datasets usually come with a fixed set of parameters, which are partly suitable for optimizing re-ID applications. In this paper, we study dataset construction parameters to improve re-ID performance. To this end, we first experiment on the temporal subsampling rate of the sequence of bounding boxes. Second, an experiment is performed on the effects of bounding-box enlargement under various temporal sampling rates. Thirdly, we analyze how the optimal choice of such dataset design parameters change with the dataset size. The experiments reveal that a performance increase of 2.1% Rank-1 is possible over a state-of-the-art re-ID model when optimizing the dataset construction parameters, thereby increasing the state-of-the-art performance from 91.9% to 94.0% Rank-1 on the popular DukeMTMC-reID dataset. The obtained results are not specific for the applied model and likely generalize to others.

X-ray Computed Tomography (CT) is an imaging modality where patients are exposed to potentially harmful ionizing radiation. To limit patient risk, reduced-dose protocols are desirable, which inherently lead to an increased noise level in the reconstructed CT scans. Consequently, noise reduction algorithms are indispensable in the reconstruction processing chain. In this paper, we propose to leverage a conditional Generative Adversarial Networks (cGAN) model, to translate CT images from low-to-routine dose. However, when aiming to produce realistic images, such generative models may alter critical image content. Therefore, we propose to employ a frequency-based separation of the input prior to applying the cGAN model, in order to limit the cGAN to high-frequency bands, while leaving low-frequency bands untouched. The results of the proposed method are compared to a state-of-the-art model within the cGAN model as well as in a single-network setting. The proposed method generates visually superior results compared to the single-network model and the cGAN model in terms of quality of texture and preservation of fine structural details. It also appeared that the PSNR, SSIM and TV metrics are less important than a careful visual evaluation of the results. The obtained results demonstrate the relevance of defining and separating the input image into desired and undesired content, rather than blindly denoising entire images. This study shows promising results for further investigation of generative models towards finding a reliable deep learning-based noise reduction algorithm for low-dose CT acquisition.

Colorectal polyps (CRP) are precursor lesions of colorectal cancer (CRC). Correct identification of CRPs during in-vivo colonoscopy is supported by the endoscopist's expertise and medical classification models. A recent developed classification model is the Blue light imaging Adenoma Serrated International Classification (BASIC) which describes the differences between non-neoplastic and neoplastic lesions acquired with blue light imaging (BLI). Computer-aided detection (CADe) and diagnosis (CADx) systems are efficient at visually assisting with medical decisions but fall short at translating decisions into relevant clinical information. The communication between machine and medical expert is of crucial importance to improve diagnosis of CRP during in-vivo procedures. In this work, the combination of a polyp image classification model and a language model is proposed to develop a CADx system that automatically generates text comparable to the human language employed by endoscopists. The developed system generates equivalent sentences as the human-reference and describes CRP images acquired with white light (WL), blue light imaging (BLI) and linked color imaging (LCI). An image feature encoder and a BERT module are employed to build the AI model and an external test set is used to evaluate the results and compute the linguistic metrics. The experimental results show the construction of complete sentences with an established metric scores of BLEU-1 = 0.67, ROUGE-L = 0.83 and METEOR = 0.50. The developed CADx system for automatic CRP image captioning facilitates future advances towards automatic reporting and may help reduce time-consuming histology assessment.

This study presents a scalable automated video surveillance framework that (1) automatically detects the occurrences of abnormal behavior patterns by both pedestrians and vehicles, and (2) directs the focus of the security personnel to the relevant camera view, thereby providing global situational awareness. Powered by deep learning, our methodology can detect both vision and location-based abnormalities, including the events of vandalism, violence, loitering, scouting, and speeding. The proposed framework requires a low initial investment cost and features both real-time detection of various abnormal behaviors and post-crime analysis in scalable form, by enabling wide-area multi-camera networks with person/object re-identification. By combining multiple functionalities in an efficient framework, the proposed system opens up new possibilities for surveillance.

Retrieving specific vehicle tracks by Natural Language (NL)-based descriptions is a convenient way to monitor vehicle movement patterns and traffic-related events. NL-based image retrieval has several applications in smart cities, traffic control, etc. In this work, we propose TIED, a text-to-image encoder-decoder model for the simultaneous extraction of visual and textual information for vehicle track retrieval. The model consists of an encoder network that enforces the two modalities into a common latent space and a decoder network that performs an inverse mapping to the text descriptions. The method exploits visual semantic attributes of a target vehicle along with a cycle-consistency loss. The proposed method employs both intra-modal and inter-modal relationships to improve retrieval performance. Our system yields competitive performance achieving the 7th position in the Natural Language-Based Vehicle Retrieval public track of the 2021 NVIDIA AI City Challenge. We demonstrate that the proposed TIED model obtains six times higher Mean Reciprocal Rank (MRR) than the baseline, achieving an MRR of 15.48. The code and models will be made publicly available.

Background and Aims Volumetric laser endomicroscopy (VLE) is an advanced imaging modality used to detect Barrett’s esophagus (BE) dysplasia. However, real-time interpretation of VLE scans is complex and time-consuming. Computer-aided detection (CAD) may help in the process of VLE image interpretation. Our aim was to train and validate a CAD algorithm for VLE-based detection of BE neoplasia. Methods The multicenter, VLE PREDICT study, prospectively enrolled 47 patients with BE. In total, 229 nondysplastic BE and 89 neoplastic (high-grade dysplasia/esophageal adenocarcinoma) targets were laser marked under VLE guidance and subsequently underwent a biopsy for histologic diagnosis. Deep convolutional neural networks were used to construct a CAD algorithm for differentiation between nondysplastic and neoplastic BE tissue. The CAD algorithm was trained on a set consisting of the first 22 patients (134 nondysplastic BE and 38 neoplastic targets) and validated on a separate test set from patients 23 to 47 (95 nondysplastic BE and 51 neoplastic targets). The performance of the algorithm was benchmarked against the performance of 10 VLE experts. Results Using the training set to construct the algorithm resulted in an accuracy of 92%, sensitivity of 95%, and specificity of 92%. When performance was assessed on the test set, accuracy, sensitivity, and specificity were 85%, 91%, and 82%, respectively. The algorithm outperformed all 10 VLE experts, who demonstrated an overall accuracy of 77%, sensitivity of 70%, and specificity of 81%. Conclusions We developed, validated, and benchmarked a VLE CAD algorithm for detection of BE neoplasia using prospectively collected and biopsy-correlated VLE targets. The algorithm detected neoplasia with high accuracy and outperformed 10 VLE experts. (The Netherlands National Trials Registry (NTR) number: NTR 6728.)

On-the-fly reconstruction of 3D indoor environments has recently become an important research field to provide situational awareness for first responders, like police and defence officers. The protocols do not allow deployment of active sensors (LiDAR, ToF, IR cameras) to prevent the danger of being exposed. Therefore, passive sensors, such as stereo cameras or moving mono sensors, are the only viable options for 3D reconstruction. At present, even the best portable stereo cameras provide an inaccurate estimation of depth images, caused by the small camera baseline. Reconstructing a complete scene from inaccurate depth images becomes then a challenging task. In this paper, we present a real-time ROS-based system for first responders that performs semantic 3D indoor reconstruction based purely on stereo camera imaging. The major components in the ROS system are depth estimation, semantic segmentation, SLAM and 3D point-cloud filtering. First, we improve the semantic segmentation by training the DeepLab V3+ model [9] with a filtered combination of several publicly available semantic segmentation datasets. Second, we propose and experiment with several noise filtering techniques on both depth images and generated point-clouds. Finally, we embed semantic information into the mapping procedure to achieve an accurate 3D floor plan. The obtained semantic reconstruction provides important clues on the inside structure of an unseen building which can be used for navigation.

Vehicle re-identification (re-ID) is based on identity matching of vehicles across non-overlapping camera views. Recently, the research on vehicle re-ID attracts increased attention, mainly due to its prominent industrial applications, such as post-crime analysis, traffic flow analysis, and wide-area vehicle tracking. However, despite the increased interest, the problem remains to be challenging. One of the most significant difficulties of vehicle re-ID is the large viewpoint variations due to non-standardized camera placements. In this study, to improve re-ID robustness against viewpoint variations while preserving algorithm efficiency, we exploit the use of vehicle orientation information. First, we analyze and benchmark various deep learning architectures in terms of performance, memory use, and cost on applicability to orientation classification. Secondly, the extracted orientation information is utilized to improve the vehicle re-ID task. For this, we propose a viewpoint-aware multi-branch network that improves the vehicle re-ID performance without increasing the forward inference time. Third, we introduce a viewpoint-aware mini-batching approach which yields improved training and higher re-ID performance. The experiments show an increase of 4.0% mAP and 4.4% rank-1 score on the popular VeRi dataset with the proposed mini-batching strategy, and overall, an increase of 2.2% mAP and 3.8% rank-1 score compared to the ResNet-50 baseline.

Convolutional Neural Networks (CNNs) are reshaping signal processing and computer vision by providing data-driven solutions for inverse problems such as noise reduction. However, their relationship with established signal processing methods is sometimes unclear and its development not fully exploiting the existing knowledge. In this paper, rather than improving existing CNNs with wavelet transformations as explored earlier, we improve the wavelet shrinkage approach to noise-reduction with a data-driven solution. The resulting CNN has clear encoding, decoding and processing paths. As application, we perform noise reduction in Dual-Energy Cone- Beam CT. The obtained results were compared to a UNet-like architecture, which reveal better noise-free images without aliasing artifacts. This indicates that that our architecture is able to preserve well the information contained in the images because the architecture exploits explicitly the underlying signal representation.

Patients suffering from Barrett's Esophagus (BE) are at an increased risk of developing esophageal adenocarcinoma and early detection is crucial for a good prognosis. To aid the endoscopists with the early detection for this preliminary stage of esophageal cancer, this work concentrates on the development and extensive evaluation of a state-of-the-art computer-aided classification and localization algorithm for dysplastic lesions in BE. To this end, we have employed a large-scale endoscopic data set, consisting of 494,355 images, in combination with a novel semi-supervised learning algorithm to pretrain several instances of the proposed neural network architecture. Next, several Barrett-specific data sets that are increasingly closer to the target domain with significantly more data compared to other related work, were used in a multi-stage transfer learning strategy. Additionally, the algorithm was evaluated on two prospectively gathered external test sets and compared against 53 medical professionals. Finally, the model was also evaluated in a live setting without interfering with the current biopsy protocol. Results from the performed experiments show that the proposed model improves on the state-of-the-art on all measured metrics. More specifically, compared to the best performing state-of-the-art model, the specificity is improved by more than 20% points while simultaneously preserving high sensitivity and reducing the false positive rate substantially. Our algorithm yields similar scores on the localization metrics, where the intersection of all experts is correctly indicated in approximately 92% of the cases. Furthermore, the live pilot study shows great performance in a clinical setting with a patient level accuracy, sensitivity, and specificity of 90%. Finally, the proposed algorithm outperforms each individual medical expert by at least 5% and the average assessor by more than 10% over all assessor groups with respect to accuracy.

Vehicle re-identification plays a crucial role in the management of transportation infrastructure and traffic flow. However, this is a challenging task due to the large view-point variations in appearance, environmental and instance-related factors. Modern systems deploy CNNs to produce unique representations from the images of each vehicle instance. Most work focuses on leveraging new losses and network architectures to improve the descriptiveness of these representations. In contrast, our work concentrates on re-ranking and embedding expansion techniques. We propose an efficient approach for combining the outputs of multiple models at various scales while exploiting tracklet and neighbor information, called dual embedding expansion (DEx). Additionally, a comparative study of several common image retrieval techniques is presented in the context of vehicle re-ID. Our system yields competitive performance in the 2020 NVIDIA AI City Challenge with promising results. We demonstrate that DEx when combined with other re-ranking techniques, can produce an even larger gain without any additional attribute labels or manual supervision.

Dual-Energy CT offers significant advantages over traditional CT imaging because it offers energy-based awareness of the image content and facilitates material discrimination in the projection domain. The Dual-Energy CT concept has intrinsic redundancy that can be used for improving image quality, by jointly exploiting the high- and low-energy projections. In this paper we focus on noise reduction. This work presents the novel noise-reduction algorithm Dual Energy Shifted Wavelet Denoising (DESWD), which renders high-quality Dual-Energy CBCT projections out of noisy ones. To do so, we first apply a Generalized Anscombe Transform, enabling us to use denoising methods proposed for Gaussian noise statistics. Second, we use a 3D transformation to denoise all the projections at once. Finally we exploit the inter-channel redundancy of the projections to create sparsity in the signal for better denoising with a channel-decorrelation step. Our simulation experiments show that DESWD performs better than a state-of-the-art denoising method (BM4D) in limited photon-count imaging, while BM4D achieves excellent results for less noisy conditions.

Visual place recognition using query and database images from different sources remains a challenging task in computer vision. Our method exploits global descriptors for efficient image matching and local descriptors for geometric verification. We present a novel, multi-scale aggregation method for local convolutional descriptors, using memory vector construction for efficient aggregation. The method enables to find preliminary set of image candidate matches and remove visually similar but erroneous candidates. We deploy the multi-scale aggregation for visual place recognition on 3 large-scale datasets. We obtain a Recall@10 larger than 94% for the Pittsburgh dataset, outperforming other popular convolutional descriptors used in image retrieval and place recognition. Additionally, we provide a comparison for these descriptors on a more challenging dataset containing query and database images obtained from different sources, achieving over 77% Recall@10.

Overweight vehicles are a common source of pavement and bridge damage. Especially mobile crane vehicles are often beyond legal per-axle weight limits, carrying their lifting blocks and ballast on the vehicle instead of on a separate trailer. To prevent road deterioration, the detection of overweight cranes is desirable for law enforcement. As the source of crane weight is visible, we propose a camera-based detection system based on convolutional neural networks. We iteratively label our dataset to vastly reduce labeling and extensively investigate the impact of image resolution, network depth and dataset size to choose optimal parameters during iterative labeling. We show that iterative labeling with intelligently chosen image resolutions and network depths can vastly improve (up to 70×) the speed at which data can be labeled, to train classification systems for practical surveillance applications. The experiments provide an estimate of the optimal amount of data required to train an effective classification system, which is valuable for classification problems in general. The proposed system achieves an AUC score of 0.985 for distinguishing cranes from other vehicles and an AUC of 0.92 and 0.77 on lifting block and ballast classification, respectively. The proposed classification system enables effective road monitoring for semi-automatic law enforcement and is attractive for rare-class extraction in general surveillance classification problems.

Surgical navigation systems can enhance surgeon vision and form a reliable image-guided tool for complex interventions as spinal surgery. The main prerequisite is successful patient tracking which implies optimal motion compensation. Nowadays, optical tracking systems can satisfy the need of detecting patient position during surgery, allowing navigation without the risk of damaging neurovascular structures. However, the spine is subject to vertebrae movements which can impact the accuracy of the system. The aim of this paper is to investigate the feasibility of a novel approach for offering a direct relationship to movements of the spinal vertebra during surgery. To this end, we detect and track patient spine features between different image views, captured by several optical cameras, for vertebrae rotation and displacement reconstruction. We analyze patient images acquired in a real surgical scenario by two gray-scale cameras, embedded in the flat-panel detector of the C-arm. Spine segmentation is performed and anatomical landmarks are designed and tracked between different views, while experimenting with several feature detection algorithms (e.g. SURF, MSER, etc.).The 3D positions for the matched features are reconstructed and the triangulation errors are computed for an accuracy assessment. The analysis of the triangulation accuracy reveals a mean error of 0.38~mm, which demonstrates the feasibility of spine tracking and strengthens the clinical application of optical imaging for spinal navigation.

In this paper, we investigate person re-identification (re-ID) in a multi-camera network for surveillance applications. To this end, we create a Spatio-Temporal Multi-Camera model (ST-MC model), which exploits statistical data on a person’s entry/exit points in the multi-camera network, to predict in which camera view a person will re-appear. The created ST-MC model is used as a novel extension to the Multiple Granularity Network (MGN) [1], which is the current state of the art in person re-ID. Compared to existing approaches that are solely based on Convolutional Neural Networks (CNNs), our approach helps to improve the re-ID performance by considering not only appearance-based features of a person from a CNN, but also contextual information. The latter serves as scene understanding information complimentary to person re-ID. Experimental results show that for the DukeMTMC-reID dataset [2][3], introduction of our ST-MC model substantially increases the mean Average Precision (mAP) and Rank-1 score from 77.2% to 84.1%, and from 88.6% to 96.2%, respectively.

The goal of endoscopic analysis is to find abnormal lesions and determine further therapy from the obtained information. For example, in case of Barrett’s esophagus, the objective of endoscopy is to timely detect dysplastic lesions, before endoscopic resection is no longer possible. However, the procedure produces a variety of non-informative frames and lesions can be missed due to poor video quality. Especially when analyzing entire endoscopic videos made by non-expert endoscopists, informative frame classification is crucial to e.g. video quality grading. This analysis involves classification problems such as polyp detection or dysplasia detection in Barrett’s Esophagus. This work concentrates on the design of an automated indication of informativeness of video frames. We propose an algorithm consisting of state-of-the-art deep learning techniques, to initialize frame-based classification, followed by a hidden Markov model to incorporate temporal information and control consistent decision making. Results from the performed experiments show that the proposed model improves on the state-of-the-art with an F1-score of 91%, and a substantial increase in sensitivity of 10%, thereby indicating improved labeling consistency. Additionally, the algorithm is capable of processing 261 frames per second, which is multiple times faster compared to other informative frame classification algorithms, thus enabling real-time computation.

Frequent pain and discomfort in premature infants can lead to long-term adverse neurodevelopmental outcomes. Video-based monitoring is considered to be a promising contactless method for identification of discomfort moments. In this study, we propose a video-based method for automated detection of infant discomfort. The method is based on analyzing facial and body motion. Therefore, motion trajectories are estimated from frame to frame using optical flow. For each video segment, we further calculate the motion acceleration rate and extract 18 time- and frequency-domain features characterizing motion patterns. A support vector machine (SVM) classifier is then applied to video sequences to recognize infant status of comfort or discomfort. The method is evaluated using 183 video segments for 11 infants from 17 heel prick events. Experimental results show an AUC of 0.94 for discomfort detection and the average accuracy of 0.86 when combining all proposed features, which is promising for clinical use.

The current paradigm in privacy protection in street-view images is to detect and blur sensitive information. In this paper, we propose a framework that is an alternative to blurring, which automatically removes and inpaints moving objects (e.g. pedestrians, vehicles) in street-view imagery. We propose a novel moving object segmentation algorithm exploiting consistencies in depth across multiple street-view images that are later combined with the results of a segmentation network. The detected moving objects are removed and inpainted with information from other views, to obtain a realistic output image such that the moving object is not visible anymore. We evaluate our results on a dataset of 1000 images to obtain a peak noise-to-signal ratio (PSNR) and L1 loss of 27.2 dB and 2.5%, respectively. To ensure the subjective quality, To assess overall quality, we also report the results of a survey conducted on 35 professionals, asked to visually inspect the images whether object removal and inpainting had taken place. The inpainting dataset will be made publicly available for scientific benchmarking purposes at https://research.cyclomedia.com, Comment: Accepted to CVPR 2019. Dataset (and provided link) will be made available before the CVPR

Colorectal polyps are an indicator of colorectal cancer (CRC). Classification of polyps during colonoscopy is still a challenge for which many medical experts have come up with visual models albeit with limited success. In this paper, a classification approach is proposed to differentiate between polyp malignancy, using features extracted from the Global Average Pooling (GAP) layer of a Convolutional Neural Network (CNNs). Two recent endoscopic modalities are used to improve the algorithm prediction: Blue Laser Imaging (BLI) and Linked Color Imaging (LCI). Furthermore, a new strategy of per-class data augmentation is adopted to tackle an unbalanced class distribution and to improve the decision of the classifiers. As a result, we increase the performance compared to state-of-the-art methods (0.97 vs 0.90 AUC). Our method for automatic polyp malignancy classification facilitates future advances towards patient safety and may avoid time-consuming and costly histopathological assessment.

Fast and accurate catheter detection in cardiac catheterization using harmless 3D ultrasound (US) can improve the efficiency and outcome of the intervention. However, the low image quality of US requires extra training for sonographers to localize the catheter. In this paper, we propose a catheter detection method based on a pre-trained VGG network, which exploits 3D information through re-organized cross-sections to segment the catheter by a shared fully convolutional network (FCN), which is called a Direction-Fused FCN (DF-FCN). Based on the segmented image of DF-FCN, the catheter can be localized by model fitting. Our experiments show that the proposed method can successfully detect an ablation catheter in a challenging ex-vivo 3D US dataset, which was collected on the porcine heart. Extensive analysis shows that the proposed method achieves a Dice score of 57.7%, which offers at least an 11.8 % improvement when compared to state-of-the-art instrument detection methods. Due to the improved segmentation performance by the DF-FCN, the catheter can be localized with an error of only 1.4 mm., ISBI 2019 accepted

Conventional raycasting methods extract 2D-images from pointclouds in two main steps. The pointcloud is voxelized and then, rays are casted from a virtual-camera center towards the model. The value for each pixel in the resulting image is calculated based on the closest non-empty voxel intersected with the corresponding ray. Both voxelizing and such raycasting limit the quality (resolution) of the extracted image and impose high memory demands. In this paper, we propose an alternative backraycasting method, where rays are casted from the model towards the virtual-camera center and intersecting an image plane. This does not require any voxel grid to be generated. Moreover, this method allows to obtain images with any required resolution with all the points involved. Besides this, a neighbours-consistency technique is introduced to enhance the resulting image quality. The proposed method has been evaluated based on several criteria and for various resolutions. Evaluation results show that the proposed method compared to the conventional approach executes upto 49 times faster and improves PSNR and SSIM metrics for the resulting images by 26% and 12%; respectively. This improvement is beneficial for such domains as feature matching, edge detection, OCR and calibration. To enable researchers generating the same results and extend this work, the dataset and implementation codes are publicly available [1].

This paper proposes a cycle-accurate performance analysis method for real-time component-based distributed systems (CB-RTDS).The method involves the following phases:(a) profiling SW components at cycle execution level and modeling the obtained performance measurements in MARTE-compatible component resource models,(b) guided composition of the system architecture from available SW and HW components,(c) automated generation of a system model, specifying both computation and network loads, and(d) performance analysis (scheduling, simulation and network analysis) of the composed system model.The method is demonstrated for a real-world case study of 3 autonomously navigating robots with advanced sensing capabilities.The case study is challenging because of the SW/HW mapping, real-time requirements and data synchronization among multiple nodes.This case-study proved that, thanks to the adopted low-level performance metrics, we are able to obtain accurate performance predictions of both computation and network delays.Moreover, the combination of analytical and simulation analysis methods enables the computation of both the guaranteed Worst Case Execution Time (WCET) and the detailed execution time-line data for real-time tasks.As a result, the analysis yields the identification of an optimal architecture, with respect to real-time deadlines, robustness and system costs.The paper main contributions are the cycle-accurate performance analysis workflow and supportive open-source ProMARTES tool-chain, both incorporating a network prediction model in all the performance analysis phases.

Instrument segmentation plays a vital role in 3D ultrasound (US) guided cardiac intervention. Efficient and accurate segmentation during the operation is highly desired since it can facilitate the operation, reduce the operational complexity, and therefore improve the outcome. Nevertheless, current image-based instrument segmentation methods are not efficient nor accurate enough for clinical usage. Lately, fully convolutional neural networks (FCNs), including 2D and 3D FCNs, have been used in different volumetric segmentation tasks. However, 2D FCN cannot exploit the 3D contextual information in the volumetric data, while 3D FCN requires high computation cost and a large amount of training data. Moreover, with limited computation resources, 3D FCN is commonly applied with a patch-based strategy, which is therefore not efficient for clinical applications. To address these, we propose a POI-FuseNet, which consists of a patch-of-interest (POI) selector and a FuseNet. The POI selector can efficiently select the interested regions containing the instrument, while FuseNet can make use of 2D and 3D FCN features to hierarchically exploit contextual information. Furthermore, we propose a hybrid loss function, which consists of a contextual loss and a class-balanced focal loss, to improve the segmentation performance of the network. With the collected challenging ex-vivo dataset on RF-ablation catheter, our method achieved a Dice score of 70.5%, superior to the state-of-the-art methods. In addition, based on the pre-trained model from ex-vivo dataset, our method can be adapted to the in-vivo dataset on guidewire and achieves a Dice score of 66.5% for a different cardiac operation. More crucially, with POI-based strategy, segmentation efficiency is reduced to around 1.3 seconds per volume, which shows the proposed method is promising for clinical use.

Guidance of needles for interventions that involve percutaneous advancing of a needle to a target inside the patient's body is one of the key uses of ultrasound, such as for biopsies, ablations, and nerve blocks. During these procedures, image-based detection of the needle can circumvent complex needle-transducer alignment by ensuring an adequate visualization of the needle during the entire procedure. However, successful localization in the sector and curvilinear transducers is challenging as the needle can be invisible or partially visible, due to the lack of received beam reflections from parts of the needle. Therefore, it is necessary to explicitly model the global information present in the data for correct localization of the needle to compensate for the lost signal. We present a novel image-based localization technique to detect partially visible needles in phased-array 3D ultrasound volumes using dilated convolutional neural networks. The proposed algorithm successfully detects the needle plane with accuracy in the submillimeter domain in the 20 measured datasets, which also consists of the cases with mostly invisible needle shaft.

By combining terrestrial panorama images and aerial imagery, or using LiDAR, large 3D point clouds can be generated for 3D city modeling. We describe an algorithm for change detection in point clouds, including three new contributions: change detection for LOD2 models compared to 3D point clouds, the application of detected changes for creating extended and textured LOD2 models, and change detection between point clouds of different years. Overall, LOD2 model-to-point-cloud changes are reliably found in practice, and the algorithm achieves a precision of 0.955 and recall of 0.983 on a synthetic dataset. Despite not having a watertight model, texturing results are visually promising, improving over directly textured LOD2 models.

For maritime surveillance, collecting information about vessels and their behavior is of vital importance. This implies reliable vessel detection and determination of the viewing angle to a vessel, which can help in analyzing the vessel behavior and in re-identification. This paper presents a vessel classification and orientation recognition system for maritime surveillance. For this purpose, we have established two novel multi-class vessel detection and vessel orientation datasets, provided to open public access. Each dataset contains 10,000 training and 1,000 evaluation images with 31,078 vessel labels (10 vessel types and 5 orientation classes). We deploy VGG/SSD to train two separate CNN models for multi-class detection and for orientation recognition of vessels. Both trained models provide a reliable F1 score of 82% and 76%, respectively.

This paper explores the use of stixels in a probabilistic stereo vision-based collision-warning system that can be part of an ADAS for intelligent vehicles. In most current systems, collision warnings are based on radar or on monocular vision using pattern recognition (and ultra-sound for park assist). Since detecting collisions is such a core functionality of intelligent vehicles, redundancy is key. Therefore, we explore the use of stereo vision for reliable collision prediction. Our algorithm consists of a Bayesian histogram filter that provides the probability of collision for multiple interception regions and angles towards the vehicle. This could additionally be fused with other sources of information in larger systems. Our algorithm builds upon the disparity Stixel World that has been developed for efficient automotive vision applications. Combined with image flow and uncertainty modeling, our system samples and propagates asteroids, which are dynamic particles that can be utilized for collision prediction. At best, our independent system detects all 31 simulated collisions (2 false warnings), while this setting generates 12 false warnings on the real-world data.

Recently, privacy has a growing importance in several domains, especially in street-view images. The conventional way to achieve this is to automatically detect and blur sensitive information from these images. However, the processing cost of blurring increases with the ever-growing resolution of images. We propose a system that is cost-effective even after increasing the resolution by a factor of 2.5. The new system utilizes depth data obtained from LiDAR to significantly reduce the search space for detection, thereby reducing the processing cost. Besides this, we test several detectors after reducing the detection space and provide an alternative solution based on state-of-the-art deep learning detectors to the existing HoG-SVM-Deep system that is faster and has a higher performance., Comment: Accepted at Electronic Imaging 2019

Change detection from ground vehicles has various applications, such as the detection of roadside Improvised Explosive Devices (IEDs). Although IEDs are hidden, they are often accompanied by visible markers, which can be any kind of object. Because of this, any suspicious change in the environment compared to an earlier moment in time, should be detected. Little work has been published to solve this ill-posed problem using deep learning. This paper shows the feasibility of applying convolutional neural networks (CNNs) to HD video, to accurately predict the presence and location of such markers in real time. The network is trained for the detection of pixel-level changes in HD video, compared to an earlier reference recording. We investigate Siamese CNNs in combination with an encoder-decoder architecture and introduce a modified double-margin contrastive loss function, to achieve pixel-level change detection results. Our dataset consists of seven pairs of challenging real-world recordings with geo-tagged test objects. The proposed network architecture is capable of comparing two images of 1920×1440 pixels in 150 ms on a GTX1080Ti GPU. The proposed network significantly outperforms state-of-the-art networks and algorithms on our dataset in terms of F-1 score, on average by 0.28.

Barrett cancer is a treatable disease when detected at an early stage. However, current screening protocols are often not effective at finding the disease early. Volumetric Laser Endomicroscopy (VLE) is a promising new imaging tool for finding dysplasia in Barrett's esophagus (BE) at an early stage, by acquiring cross-sectional images of the microscopic structure of BE up to 3-mm deep. However, interpretation of VLE scans is difficult for medical doctors due to both the size and subtlety of the gray-scale data. Therefore, algorithms that can accurately find cancerous regions are very valuable for the interpretation of VLE data. In this study, we propose a fully-automatic multi-step Computer-Aided Detection (CAD) algorithm that optimally leverages the effectiveness of deep learning strategies by encoding the principal dimension in VLE data. Additionally, we show that combining the encoded dimensions with conventional machine learning techniques further improves results while maintaining interpretability. Furthermore, we train and validate our algorithm on a new histopathologically validated set of in-vivo VLE snapshots. Additionally, an independent test set is used to assess the performance of the model. Finally, we compare the performance of our algorithm against previous state-of-the-art systems. With the encoded principal dimension, we obtain an Area Under the Curve (AUC) and F1 score of 0.93 and 87.4% on the test set respectively. We show this is a significant improvement compared to the state-of-the-art of 0.89 and 83.1%, respectively, thereby demonstrating the effectiveness of our approach.

Object detection and classification of traffic signs in street-view imagery is an essential element for asset management, map making and autonomous driving. However, some traffic signs occur rarely and consequently, they are difficult to recognize automatically. To improve the detection and classification rates, we propose to generate images of traffic signs, which are then used to train a detector/classifier. In this research, we present an end-to-end framework that generates a realistic image of a traffic sign from a given image of a traffic sign and a pictogram of the target class. We propose a residual attention mechanism with dense concatenation called Dense Residual Attention, that preserves the background information while transferring the object information. We also propose to utilize multi-scale discriminators, so that the smaller scales of the output guide the higher resolution output. We have performed detection and classification tests across a large number of traffic sign classes, by training the detector using the combination of real and generated data. The newly trained model reduces the number of false positives by 1.2 - 1.5% at 99% recall in the detection tests and an absolute improvement of 4.65% (top-1 accuracy) in the classification tests., The first two authors have equal contribution. Accepted at International Conference on Pattern Recognition 2018 (ICPR)

The incidence of Esophageal Adenocarcinoma (EAC), a form of esophageal cancer, has rapidly increased in recent years. Dysplastic tissue can be removed endoscopically at an early stage, and since survival chances of patients are limited at later stages of the disease, early detection is of key impor- tance. Recently, several CAD systems for HD endoscopic images have been proposed, but these are computationally expensive, making them unfit for clinical use requiring real- time analysis. In this paper, we present a novel approach for early esophageal cancer detection using Transfer Learning with CNNs. Given the small amount of annotated data, CNN Codes are applied, where intermediate layers of the net- work are used as features for conventional classifiers. Various classifiers are combined with four of the most widely-used networks. Additionally, sliding windows are used to obtain a coarse-grained annotation indicating any possible cancerous regions. This approach outperforms the current state-of-the-art with a frame-based AUC of 0.92, while allowing both near real-time prediction and annotation at 2 fps, in a MATLAB-based framework.

3D Ultrasound (US) image-based catheter detection can potentially decrease the cost on extra equipment and training. Meanwhile, accurate catheter detection enables to decrease the operation duration and improves its outcome. In this paper, we propose a catheter detection method based on convolutional neural networks (CNNs) in 3D US. Voxels in US images are classified as catheter (or not) using triplanar-based CNNs. Our proposed CNN employs two-stage training with weighted loss function, which can cope with highly imbalanced training data and improves classification accuracy. When compared to state-of-the-art handcrafted features on ex-vivo datasets, our proposed method improves the F2-score with at least 31%. Based on classified volumes, the catheters are localized with an average position error of smaller than 3 voxels in the examined datasets, indicating that catheters are always detected in noisy and low-resolution images.

This paper presents a new algorithm for dual-camera 3D head tracking, intended for clinical infant monitoring. The paper includes a brief motivation with reference to the state-of-the-art in face-related image analysis. The proposed algorithm uses a clipped-ellipsoid head model and 3D head pose recovery by joint alignment of paired templates based on dense-HOG features. In the algorithm, template pairs are dynamically extracted and a limited number of template pairs are stored and re-used for drift reduction. We report experimental results on real-life videos of infants in bed in a hospital, captured in visual light as well as near-infrared light. Results show consistently good tracking behavior. For challenging video sequences, the mean tracking error in terms of endocanthion location error relative to the innercanthal distance remains below 30%. This error has proven to be sufficiently low for 3D head tracking to support infant face analysis. For this reason, the proposed algorithm is used successfully in an infant monitoring system under development.

In this paper we describe and verify a method, called SMIP, to circumvent the trade-off between motion blur and noise, specifically for scenes with predominantly two distinct linear motions (sparse motion). This is based on employing image stabilization hardware to track objects during exposure while capturing two images in quick succession. The two images are combined into a single sharp image without segmentation or local motion estimation. We provide a theoretical analysis and simulations to show that the Signal-to-Noise Ratio (SNR) increases up to 20 dB over conventional short-exposure photography. We demonstrate that the proposed method significantly improves the SNR compared to existing methods. Furthermore, we evaluate a proof-of-concept using modified off-the-shelf optical image stabilization hardware to verify the effectiveness of our method in practice, showing a good correspondence between the simulation and practical results.