Search

Skin cancer is one of the most common types of cancer in Brazil and its incidence rate has increased in recent years. Melanoma cases are more aggressive compared to nonmelanoma skin cancer. Machine learning-based classification algorithms can help dermatologists to diagnose whether skin lesion is melanoma or non-melanoma cancer. We compared four convolutional neural networks architectures (ResNet-50, VGG16, Inception-v3, and DenseNet-121) using different training strategies and validation methods to classify seven classes of skin lesions. The experiments were executed using the HAM10000 dataset which contains 10,015 images of pigmented skin lesions. We considered the test accuracy to determine the best model for each strategy. DenseNet-121 was the best model when trained with fine-tuning and data augmentation, 90% (k-fold crossvalidation). Our results can help to improve the use of machine learning algorithms for classifying pigmented skin lesions.

The differentiation between the charcoal produced from (Eucalyptus) plantations and native forests is essential to control, commercialization, and supervision of its production in Brazil. The main contribution of this study is to identify the charcoal origin using macroscopic images and Deep Learning Algorithm. We applied a Convolutional Neural Network (CNN) using VGG-16 architecture, with preprocessing based on contrast enhancement and data augmentation with rotation over the training set images. on the performance of the CNN with fine-tuning using 360 macroscopic charcoal images from the plantation and native forests. The results pointed out that our method provides new perspectives to identify the charcoal origin, achieving results upper 95 % of mean accuracy to classify charcoal from native forests for all compared preprocessing strategies.

Data augmentation is a key procedure in many image classification tasks, mainly to overcome the problem of small datasets. In this work, we exploit the data augmentation as a hyperparameter optimization approach. We tested our methods to classify erythrocytes to assist the diagnosis of sickle cell anemia. In this study, we proposed a data augmentation approach based on hyperparameter optimization to find the best augmentation policies through the Bayesian optimization algorithm. We also developed a convolutional neural network architecture from scratch and compared it with two classic architectures to classify sickle cell images. Our approach defines the best data augmentation solutions and sends those solutions to be carried out by CNN in the final training. All experiments were validated using a stratified five-fold cross-validation procedure, and our best result achieves 92.54% of accuracy. The results suggest the best augmentation policies defined with optimization allow us to obtain superior results than other strategies such as without data augmentation, several randomly defined image transformations, and only random rotations. As far as we know, our paper is the first to propose optimizing data augmentation policies in biomedical images leading to a better exploration of these strategies in several fields.

The chromosome segmentation is the most important step in automatic karyotype assembling. In this work, we presented a brand new chromosome image dataset and proposed methods for segmenting the chromosomes. Chromosome images are usually low quality, especially fish chromosomes. In order to overcome this issue, we tested three filters to reduce noise and improve image quality. After filtering, we applied adaptive threshold segmentation combined with mathematical morphology and supervised classification methods. Support Vector Machine and k-nearest neighbors were applied to discriminate between chromosomes and image background. The proposed method was applied to segment chromosomes in a new dataset. To enable measure the performance of the methods all chromosomes were manually delineated. The results are evaluated considering the Hausdorff distance and normalized sum of distances between segmented and reference images.

Coronavirus Disease 2019 (COVID-19) pandemic rapidly spread globally, impacting the lives of billions of people. The effective screening of infected patients is a critical step to struggle with COVID-19, and treating the patients avoiding this quickly disease spread. The need for automated and scalable methods has increased due to the unavailability of accurate automated toolkits. Recent researches using chest X-ray images suggest they include relevant information about the COVID-19 virus. Hence, applying machine learning techniques combined with radiological imaging promises to identify this disease accurately. It is straightforward to collect these images once it is spreadly shared and analyzed in the world. This paper presents a method for automatic COVID-19 detection using chest Xray images through four convolutional neural networks, namely: AlexNet, VGG-11, SqueezeNet, and DenseNet-121. This method had been providing accurate diagnostics for positive or negative COVID-19 classification. We validate our experiments using a ten-fold cross-validation procedure over the training and test sets. Our findings include the shallow fine-tuning and data augmentation strategies that can assist in dealing with the low number of positive COVID-19 images publicly available. The accuracy for all CNNs is higher than 97.00%, and the SqueezeNet model achieved the best result with 99.20%.

Currently, the use of unmanned aerial vehicles (UAVs) is becoming ever more common for acquiring images in precision agriculture, either to identify characteristics of interest or to estimate plantations. However, despite this growth, their processing usually requires specialized techniques and software. During flight, UAVs may undergo some variations, such as wind interference and small altitude variations, which directly influence the captured images. In order to address this problem, we proposed a Convolutional Neural Network (CNN) architecture for the classification of three linear distortions common in UAV flight: rotation, translation and perspective transformations. To train and test our CNN, we used two mosaics that were divided into smaller individual images and then artificially distorted. Results demonstrate the potential of CNNs for solving possible distortions caused in the images during UAV flight. Therefore this becomes a promising area of exploration.

Maize is an important food crop in the world, but several diseases affect the quality and quantity of agricultural production. Identifying these diseases is a very subjective and time-consuming task. The use of computer vision techniques allows automatizing this task and is essential in agricultural applications. In this study, we assess the performance of three state-of-the-art convolutional neural network architectures to classify maize leaf diseases. We apply enhancement methods such as Bayesian hyperparameter optimization, data augmentation, and fine-tuning strategies. We evaluate these CNNs on the maize leaf images from PlantVillage dataset, and all experiments were validated using a five-fold cross-validation procedure over the training and test sets. Our findings include the correlation between the maize leaf classes and the impact of data augmentation in pre-trained models. The results show that maize leaf disease classification reached 97% of accuracy for all CNNs models evaluated. Also, our approach provides new perspectives for the identification of leaf diseases based on computer vision strategies.

The topological mapping of Multielectrode Array (MEA) in vitro cultured neurons is useful in neuroscience. In this context, it is proposed a framework for the visualization and grouping of the neurons, through their electrophysiogical activities, identifying spikes and bursts, in the MEA signal recordings. The developed framework is applied to the recording signals of the Experiment 371, of dissociated in vitro neuron culture conducted at the Laboratory of Neuroengineering and Bionanotechnology Group, University of Genoa, Italy. The main motivation of this work, is the use of an estimated map of population of neurons, in an arbitrary moment of the culture to study their dynamics. It is emphasized that the visualization of the map of neurons on MEA without the need of microscopy can be a great advantage, since a photography on the culture can be not available at a specific moment of signal recording. A mathematical model of neurons proposed by Izhikevich to simulate the action potentials of in vitro cultures is used to validate the proposal.

Autoimmune diseases are the third highest cause of mortality in the world, and the identification of an anti-nuclear antibody via an immunofluorescence test for HEp-2 cells is a standard procedure to support diagnosis. In this work, we assess the performance of six preprocessing strategies and five state-of-the-art convolutional neural network architectures for the classification of HEp-2 cells. We also evaluate enhancement methods such as hyperparameter optimization, data augmentation, and fine-tuning training strategies. All experiments were validated using a five-fold cross-validation procedure over the training and test sets. In terms of accuracy, the best result was achieved by training the Inception-V3 model from scratch, without preprocessing and using data augmentation (98.28%). The results suggest the conclusions that most CNNs perform better on non-preprocessed images when trained from scratch on the analyzed dataset, and that data augmentation can improve the results from all models. Although fine-tuning training did not improve the accuracy compared to training the CNNs from scratch, it successfully reduced the training time.

Autoimmune diseases are the third cause of mortality in the world. The identification of anti-nuclear antibody (ANA) via Immunofluorescence (IIF) test in human epithelial type-2 cells (HEp-2) is a conventional method to support the diagnosis of such diseases. In the present work, three popular Convolutional Neural Networks (CNNs) are evaluated for this task: LeNet-5, AlexNet, and GoogLeNet. We also assess the impact of six different pre-processing strategies on the performance of these CNNs. Additionally, data augmentation based on the rotation of the training set images after the pre-processing strategies was evaluated. Our work is the first to consider AlexNet and GoogLeNet models for the proposed analysis and classification of HEp-2 cells images, besides the LeNet-5. Experimental results allow to conclude that neither pre-processing strategies were essential to improve accuracy values of the CNNs. However, when data augmentation is considered, contrast enhancement followed by data centralization is significant in order to achieve good results. Additionally, our results were compared with results from other state-of-art papers. Our best results were achieved by GoogLeNet architecture trained with images with no pre-processing and no data augmentation, resulting in 98.17% of accuracy, which outperforms the results presented in other works in literature.

Recently we have witnessed research efforts into developing real-time hybrid systems implementing interactions between computational models and live tissues, in an attempt to learn more about the functioning of biological neural networks. A fundamental role in the development of such systems is played by Multi-Electrode Array MEA. In vitro cultures of neurons on MEAs, however, have some drawbacks such as: needing a rigorous adherence to sterile techniques, careful choice and replenishment of media and maintenance of pH, temperature, and osmolarity. An alternative way to study and investigate live tissues which partially circumvent some of the problems with in vitro cultures is by simulating them. This paper describes the proposal of Sim-MEA, a system for modeling and simulating neuron's communications in a MEA-based in vitro culture. Sim-MEA implements a modified Izhikevich model that takes into account both: distances between neurons and distances between microelectrodes and neurons. The system also provides ways of simulating microelectrodes and their recorded signals as well as recovering experimental MEA culture data, from their images. The soundness of the Sim-MEA simulation procedure was empirically evaluated using data from an experimental in vitro cultured hippocampal neurons of Wistar rat embryos. Results from simulations, compared to those of the in vitro experiment, are presented and discussed. The paper also describes a few experiments varying several parameter values to illustrate and detail the approach.

This work presents an implementation of neocognitron neural network, using a high performance computing architecture based on GPU (graphics processing unit). Neocognitron is an artificial neural network, proposed by Fukushima and collaborators, constituted of several hierarchical stages of neuron layers, organized in two-dimensional matrices called cellular planes. For the high performance computation of face recognition application using neocognitron it was used CUDA (compute unified device architecture) as API (application programming interface) between the CPU and the GPU, from GeForce 8800 GTX of NVIDIA company, with 128 ALU's. As face image databases it was used a face database created at UFSCar, and the CMU-PIE (Carnegie Mellon University pose, illumination and expression) database. The load balancing was achieved through the use of cellular connections as threads organized in blocks, following the CUDA philosophy of development. The results showed the feasibility of this type of device as a massively parallel data processing tool, and that smaller the granularity and the data dependency of the parallel processing, better is its performance.

The reconstruction of a three-dimensional surface from a set of unorganized points is a fundamental process in a lot of applications, including laser range scanners, medical imaging, and others. This work presents a modified version of the Neural Meshes algorithm for surface reconstruction from point cloud, called Neural Meshes with Edge Swap, or Neural Meshes ES. From the original basic Neural Meshes algorithm we developed a new heuristic that include an edge-swap operation, making the algorithm less sensible to parameters variation, improving the quality of the generated surfaces. The results show that the edge-swap operation is able to avoid a series of incorrect reconstructed areas in the mesh, mainly in the concave structures.

This ongoing project describes neural network applications for helping musical composition using as inspiration the natural landscape contours. We propose supervised and unsupervised learning approaches, by using Back-Propagation-Through-Time (BPTT) and Self Organizing Maps (SOM) neural networks. In the supervised learning, the network learns certain aspects of musical structure by means of measure examples taken from melodies of the training set and uses these measures learned to compose new melodies using as input the extracted data of the landscapes contour. In the unsupervised learning, the network also uses measure examples as input during training and the extracted data of the landscapes contour in the composition stage. The obtained results show the viability of both approaches.

Microelectrode Arrays (MEA) are devices for long term electrophysiological recording of extracellular spontaneous or evocated activities on in vitro neuron culture. This work proposes and develops a framework for quantitative and morphological analysis of neuron cultures on MEAs, by processing their corresponding images, acquired by fluorescence microscopy. The neurons are segmented from the fluorescence channel images using a combination of segmentation by thresholding, watershed transform, and object classification. The positioning of microelectrodes is obtained from the transmitted light channel images using the circular Hough transform. The proposed method was applied to images of dissociated culture of rat dorsal root ganglion (DRG) neuronal cells. The morphological and topological quantitative analysis carried out produced information regarding the state of culture, such as population count, neuron-to-neuron and neuron-to-microelectrode distances, soma morphologies, neuron sizes, neuron and microelectrode spatial distributions. Most of the analysis of microscopy images taken from neuronal cultures on MEA only consider simple qualitative analysis. Also, the proposed framework aims to standardize the image processing and to compute quantitative useful measures for integrated image-signal studies and further computational simulations. As results show, the implemented microelectrode identification method is robust and so are the implemented neuron segmentation and classification one (with a correct segmentation rate up to 84%). The quantitative information retrieved by the method is highly relevant to assist the integrated signal-image study of recorded electrophysiological signals as well as the physical aspects of the neuron culture on MEA. Although the experiments deal with DRG cell images, cortical and hippocampal cell images could also be processed with small adjustments in the image processing parameter estimation.

Deep learning é uma técnica do ramo de aprendizado de máquina que tem obtido grandes resultados em diversas tarefas, quando comparada às demais técnicas da área. Dentre tais tarefas, detecção e classificação de objetos em imagens destacam-se como exemplos notáveis de sucesso. Normalmente, nesse tipo de aplicação, uma única rede neural convolucional realiza o processo de detecção das regiões de interesse, delimitando assim a área que contém o objeto a ser identificado, bem como a classificação dessa região em uma classe. A fim de melhorar os resultados na detecção e classificação de áreas de doenças em imagens de folhas de plantas, este projeto tem como objetivo investigar uma abordagem que utiliza redes neurais convolucionais compostas por dois estágios independentes, um para realizar a detecção e outro a classificação das referidas regiões de doenças, por meio de aprendizado supervisionado. Para validar a abordagem, foram realizados experimentos com três diferentes conjuntos de dados compostos por imagens de folhas de macieira, afetadas por doenças como ferrugem e sarna, com a tarefa de detectar e classificar as regiões de doenças. Os resultados indicam que a abordagem de dois estágios tende a melhorar a precisão média da detecção em imagens de diferentes conjuntos de dados, além de permitir uma melhor transferência de aprendizado quando conjuntos de dados não vistos são usados para teste. Esta abordagem também permite maior flexibilidade na escolha de redes de detecção e classificação para adequar o modelo a cenários específicos. Além disso, as visualizações dos mapas de características dos modelos indicam que as redes de dois estágios apresentam mapas com regiões de ativação mais acentuadas, facilitando a interpretação dos resultados. Por fim, este trabalho também mostrou ser possível a utilização de tais redes neurais por meio de um protótipo de aplicativo para dispositivos móveis (como smartphones e tablets), permitindo um diagnóstico instantâneo das doenças e a criação de uma base colaborativa de novas imagens, bem como difundindo o uso da tecnologia pela população em geral. Deep learning is a technique in the machine learning branch that has achieved great results in several tasks when compared to other techniques in the area. Among such tasks, object detection and classification of images stand out as notable examples of success. Normally, in this type of application, a single convolutional neural network performs the process of detecting regions of interest, thus delimiting the area that contains the object to be identified, as well as the classification of that region. In order to improve results in the detection and classification of disease areas in plant leaf images, this project aims to investigate an approach that employs convolutional neural networks composed of two independent stages, one to perform the detection and the other to perform the classification of regions of diseases, by using supervised learning. In order to validate the approach, experiments were carried out using different datasets composed of images of apple leaves, affected by diseases such as rust and scab, with the task of detecting and classifying disease regions. Results indicate that the two-stage approach tends to improve the mean average precision (mAP) of detection on images from different datasets, in addition to allowing better transfer of learning when unseen datasets are used for testing. This approach also allows greater flexibility in choosing the detection and classification networks to tailor the model to specific scenarios. In addition, the visualizations of the feature maps yielded by the models indicate that the two-stage networks have maps with more accentuated activation regions, facilitating the interpretation of the results. Finally, this work also showed that it is possible to use such neural networks through an application prototype for mobile devices (such as smartphones and tablets), allowing an instant diagnosis of plant diseases as well as the creation of a collaborative dataset of new images, spreading the use of technology by the general population.