The number of people diagnosed with autism spectrum disorder (ASD) has been increasing significantly. However, the early treatment has become restricted, reducing the potential benefits from this intervention and increasing family and social expenses as a result of late interventions. Underdeveloped countries have proportionally fewer people diagnosed with ASD. Thus, the accessibility to mechanisms used to identification of ASD is important in theses countries. The aim of this paper is to evaluate the use of convolutional neural networks to assist in the identification of ASD, using static two-dimensional facial images as input. The results show success for the studied approach. MobileNet and DenseNet201 obtained the best results with an average of 90.7% accuracy and standard deviations of 0.68% and 1.64%, respectively. DenseNet201 achieved an accuracy of 93.5% in the best cases.

In a small town, a funeral is held after the death of a gardener. Three brothers find themselves lost in their imaginations on the day of the burial.

Diabetic Foot Ulcers (DFU) are lesions in the foot region caused by diabetes mellitus. It is essential to define the appropriate treatment in the early stages of the disease once late treatment may result in amputation. This article proposes an ensemble approach composed of five modified convolutional neural networks (CNNs) - VGG-16, VGG-19, Resnet-50, InceptionV3, and Densenet-201 - to classify DFU images. To define the parameters, we fine-tuned the CNNs, evaluated different configurations of fully connected layers, and used batch normalization and dropout operations. The modified CNNs were well suited to the problem; however, we observed that the union of the five CNNs significantly increased the success rates. We performed tests using 8,250 images with different resolution, contrast, color, and texture characteristics and included data augmentation operations to expand the training dataset. 5-fold cross-validation led to an average accuracy of 95.04%, resulting in a Kappa index greater than 91.85%, considered “Excellent”.

Dermoscopic images are commonly used in the early diagnosis of skin lesions, and several computational systems have been proposed to analyze them. The segmentation of the lesions is a fundamental step in many of these systems. Therefore, a semi-automatic segmentation method is proposed here, which begins by building the superpixels of the image under analysis based on the zero parameter version of the simple linear iterative clustering (SLIC0) algorithm. Then, each superpixel is represented using a descriptor built by combining the grey-level co-occurrence matrix and Tamura texture features. Afterward, the gain ratios of the features are used to select the input for the semi-supervised seeded fuzzy C-means clustering algorithm. Hence, from a few specialist-selected superpixels, this clustering algorithm groups the built superpixels into lesion or background regions. Finally, the segmented image undergoes a post-processing step to eliminate sharp edges. The experiments were performed on 1380 images: 401 images from the PH2 and DermIS datasets, which were used to establish the parameters of the method, and 3,573 images from the ISIC 2016, ISIC 2017 and ISIC 2018 datasets were used for the analysis of the method’s performance. The findings suggest that, by manually identifying just a few of the generated superpixels, the method can achieve an average segmentation accuracy of 96.78%, which confirms its superiority to the ones in the literature.

Melanoma is considered the worst type of skin cancer. The early diagnosis of this disease is still a complex task due to many variables that must be analyzed. Because of this, new methodologies are becoming common in the literature due to the good results obtained. Convolutional Neural Networks are Deep Learning techniques capable of providing effective solutions in the classification of medical images. In this sense, this work developed a disease detection system using AlexNet and VGG-F convolutional architectures, trained with images of skin lesions to create feature descriptors, not classifiers. Other conventional descriptors of skin lesions were used to assess the quality of data obtained from the last layers of convolutional architectures. Data from all feature extraction processes were submitted to the conventional classifiers Support Vector Machine, Multilayer Perceptron, and K-Nearest Neighbor. The results obtained in the approach show that the feature extracting models are viable and can offer a more accurate melanoma diagnosis possibility. The VGG-F architecture obtained the best result, with an accuracy of 91.54% and a precision of 91.64% given by the K-Nearest Neighbor. It is possible to see that this result highlights the quality of data in convolutional architectures and can provide a sense of further research.

The analysis of pollen grains is a prominent task in areas such as ecology, food engineering, and others that have different purposes, such as identifying the origin of honey, as well as helping in the development of new products or evaluating the quality of the products. This research presents a CNN architecture to classify pollen grains that can have performance equal to or superior to those found in the literature. Using POLEN23E database. Two experiments were performed with this database, one of which used data augmentation to improve accuracy. Promising results were obtained, as the experiments achieved 92% accuracy in the worst case and 100% accuracy in the best case. Two experiments were performed where one of them used data augmentation to improve accuracy. Promising results were obtained, as the experiments achieved 92% accuracy in the worst case and 100% accuracy in the best case.

Lung cancer is the most prevalent cancer in the world and early detection and diagnosis enable more treatment options and a far greater chance of survival. In this work, we propose an algorithm based on 3D Convolutional Neural Network (CNN) to classify pulmonary nodules as benign or malignant in computed tomography images. Three architecture of 3D CNNs are proposed, containing different input sizes and numbers of convolutional layers. In addition, we investigated data augmentation techniques and modifications in the network training cost function to address the problem of imbalanced data. The best result was achieved for input size of 32×32×32 pixels, 2 blocks of convolutional layers and 2 pooling layers. Also, the modification of cost function achieved promising results, with accuracy of 0.9188, kappa of 0.8019, sensitivity of 0.8481, specificity of 0.9479 and AUC of 0.8980 in the test set during malignant nodule detection.

Clustering is one of the main pattern recognition techniques. This technique consists in organizing the elements into groups (clusters) taking into consideration some metric that allows determining the similarity of them. These data sets often describe the elements that compose them through attributes that can assume various values types, requiring efficient methods in the detecting correlations task between complex (or mixed) type data. However, the clustering process does not provide clear information that allows inferring the characteristics of each cluster formed, that is, the clustering process result does not allow the clusters to have their meaning easily understood. Data labeling aims at identifying these characteristics and then allowing full understanding of the resulting clusters. This paper proposes the joint use of unsupervised and supervised Machine Learning methods for data clustering and labeling tasks, respectively. The labeling task consists in identifying the clusters through their most relevant characteristics. The algorithms used are known to be efficient, obtaining satisfactory results in the definitions of the clusters formed in the experiments exposed here.

One of the least invasive means of diagnosing certain diseases is through the use of imaging tests. Automated analysis usually begins by detecting the regions to be investigated and extracting relevant information such as shape and texture. In this paper, we propose a semi-supervised clustering algorithm called Seeded Fuzzy C-means to segment regions of interest in medical images. Seeded Fuzzy C-means is based on the well-known Fuzzy C-means; however, it also uses information provided by the physician to insert constraints during group choices. In this way, the element is placed in a group with a higher degree of certainty. We evaluated the proposed algorithm on a total of 2,200 images in leukemia, skin cancer, cervical cancer, and glaucoma image databases. Two semi-supervised clustering algorithms were implemented from literature to perform this analysis. The results illustrate the ability of the algorithm to efficiently assist physicians in detecting regions of interest, as it achieved an “Excellent” Kappa index in most of the tests.

Melanoma is a malignant skin lesion, and it is currently among the most dangerous existing cancers. However, early diagnosis of this disease gives the patient a higher chance of cure. In this work, a computational method was designed to assist dermatologists in the diagnosis of skin lesions as melanoma or non-melanoma using dermoscopic images. We conducted an extensive study to define the best set of attributes for image representation. In total, we evaluated 12,705 characteristics and three classifiers. The proposed approach aims to classify skin lesions using a hybrid descriptor obtained by combining features of color, shape, texture and pre-trained Convolutional Neural Networks. These characteristics are used as inputs to a MultiLayer Perceptron classifier. The results are promising, reaching an accuracy of 92.1% and a Kappa index of 0.8346 in 406 images from two public image databases.

Glaucoma is a disease that affects the optic nerve, making it difficult to transmit impulses from the retina to the brain. This failure is caused by increased intraocular pressure and can lead to blindness. Early detection is the best alternative to mitigate the effects of glaucoma since the disease has no cure. This paper proposes a semi-automatic method to calculate the ratio between the diameters of the optic disc and cup. The approach begins repainting the blood vessels, minimizing their interference in the next steps. We perform the segmentation in two stages. Initially, we used the SLIC algorithm for superpixels generation. Subsequently, the algorithm Seeded Fuzzy C-means groups the superpixels based on a small percentage of data labeled by specialists. Then the images undergo a post-processing step eliminating noise and smoothing edges. This last phase aims to adapt the segmented images for the CDR calculation. We evaluated the method in 260 images from two public databases, DRISHTI-GS (101) and RIM-ONE r3 (159). The analysis reached an accuracy in the segmentation above 91% and a mean absolute error of 0.0779 in the CDR calculation.

Glaucoma is an eye disease that causes irreversible vision loss. Retinography is done manually by the ophthalmologist and is the cheapest, least invasive and most effective way to diagnose glaucoma. The ratio between the diameter of the outer part of the Optic Disc (OD) and the cup (internal part) called CDR (cup-to-disc ratio) is an important indicator of glaucoma presence in patients. This paper proposes a semiautomatic approach that includes the segmentation of OD and cup regions. The proposed approach consists of four stages. The first stage consists of preprocessing the retinal image, in order to remove blood vessels and a possible influence in the segmentation stage. In the second stage we apply the Seeded Fuzzy C-means algorithm to segment the preprocessed image in order to indentify cup and OD. The third step involves the application of a post-processing so that non-segmented regions are filled. Finally, the last step calculates the value of the CDR associated with the retinal image. To verify the applicability of the proposed approach, we carried out tests in two public image databases: DRISHTI-GS and RIM-ONE r3. The results obtained illustrate the feasibility of applying the approach in order to effectively assist ophthalmologists in the segmentation of cup and OD, as well as the calculation of the CDR.

Leukemia is a type of cancer that originates in the bone marrow and is characterized by abnormal proliferation of white blood cells. In order to have correct identification of lymphoblasts, hematologists examine blood blades of the patient. A low cost and efficient solution to facilitate the work of these experts is the use of systems to examine blood microscopic images. Segmentation is considered a crucial step to developing these systems. In this paper, we propose an automatic segmentation technique that uses two-color systems and the clustering algorithm K-means. The proposed approach is evaluated on three public image databases with different characteristics and performance measures used are: accuracy, specificity, sensitivity and Kappa index. The results obtained in the experiments have Kappa index of 0.9306 in ALL-IDB 2, 0.8603 in BloodSeg and 0.9119 in Leukocytes database. These measures outperform other methods of literature.

Motorcycle accidents have been rapidly growing throughout the years in many countries. Due to various social and economic factors, this type of vehicle is becoming increasingly popular. The helmet is the main safety equipment of motorcyclists, but many drivers do not use it. If an motorcyclist is without helmet an accident can be fatal. This paper aims to explain and illustrate an automatic method for motorcycles detection and classification on public roads and a system for automatic detection of motorcyclists without helmet. For this, a hybrid descriptor for features extraction is proposed based in Local Binary Pattern, Histograms of Oriented Gradients and the Hough Transform descriptors. Traffic images captured by cameras were used. The best result obtained from classification was an accuracy rate of 0.9767, and the best result obtained from helmet detection was an accuracy rate of 0.9423.

In recent years have increased the use of automated mechanisms for monitoring and enforcement of fines for traffic violations, such as radar, electronic spines and photosensors. Due to various economic and social factors use of motorcycles is gaining increasingly popular acceptance. Increasing the number of such vehicle with carelessness by conductors made to grow abruptly the number of accidents. The main security equipment of motorcyclists is the helmet but many conductors do not use it or use incorrectly. This work aims to study and implement methods for automatic detection of motorcyclists on public roads in order to, in a future work, the detection of non-use of helmets. For this purpose, transit images were used captured by video cameras. From these images different attributes have been taken through the SURF, HAAR, HOG, Fourier and K-means descriptors. And for classification of images were used classifiers as Multilayer Perceptron and Support Vector Machine.

Microaneurysms (MAs) are the earliest sign of diabetic retinopathy and manifest as small reddish spots on the retina. Generally, algorithm design for MAs detection starts by separating the vascular system from the background for a posterior analysis of candidate MAs presence. Following this approach, this paper assesses three different methods for vessel segmentation and how they affect posterior MAs detection. The robustness in developing automatic screening systems for MAs detection is discussed and a methodology to detect candidate MAs in retinal images is introduced. The algorithm combines different vessel segmentation methods with region growing to evaluate which is the best to provide candidate MAs detection.