Comparing Prediction of Early TBI Mortality with Multilayer Perceptron Neural Network and Convolutional Neural Network.

In this work, we compare the performance of a multilayer perceptron neural network and convolutional networks for the prediction of 14-day mortality in patients with TBI, using a database obtained in a low-and middle-income country, with 529 records and 16 predictor variables. The missing values of several variables were filled in with techniques such as decision tree, random forest, k-nearest-neighbor and linear regression. In the simulation of neural networks, several optimization methods were used, such as RMSProp, Adam, Adamax and SGDM. The best results obtained for the prediction rate were an accuracy of 0.845 and an area under the ROC curve of 0.911. Clinical Relevance- This proposes the prediction of early mortality in patients with TBI with an area under ROC curve of 0.911.

A New Approach to Classify Cardiac Arrythmias Using 2D Convolutional Neural Networks.

Cardiovascular diseases are the number one cause of death worldwide. Detecting cardiovascular diseases in its early stages could effectively reduce the mortality rate by providing timely treatment. In this study, we propose a new methodology to detect arrythmias, using 2D Convolutional Neural Networks. The main characteristic of the proposed methodology is the use of 15 x15 pixels gray-level images, containing the values of a heartbeat of the ECG signal. This work aims to detect 17 arrythmias. To validate and test the proposed methodology, MIT-BIH database, the main benchmark database available in literature, was used. When compared to other results previously published, the obtained precision, 92.31%, is in the state-of-the-art.Clinical Relevance- The presented work provides an automatic method to detect arrythmias in ECG signals by a new methodology.

Lung Region Segmentation in Chest X-Ray Images using Deep Convolutional Neural Networks.

Lung cancer is, by far, the leading cause of cancer death in the world. Tools for automated medical imaging analysis development of a Computer-Aided Diagnosis method comprises several tasks. In general, the first one is the segmentation of region of interest, for example, lung region segmentation from Chest X-ray imaging in the task of detecting lung cancer. Deep Convolutional Neural Networks (DCNN) have shown promising results in the task of segmentation in medical images. In this paper, to implement the lung region segmentation task on chest X-ray images, was evaluated three different DCNN architectures in association with different regularization (Dropout, L2, and Dropout + L2) and optimization methods (SGDM, RMSPROP and ADAM). All networks were applied in the Japanese Society of Radiological Technology (JSRT) database. The best results were obtained using Dropout + L2 as regularization method and ADAM as optimization method. Considering the Jaccard Coefficient obtained (0.97967 ± 0.00232) the proposal outperforms the state of the art.Clinical Relevance- The presented method reduces the time that a professional takes to perform lung segmentation, improving the effectiveness.

Automatic Bacillus Detection in Light Field Microscopy Images Using Convolutional Neural Networks and Mosaic Imaging Approach.

Tuberculosis (TB) is one of the top 10 causes of death worldwide. The diagnosis and treatment of TB in its early stages is fundamental to reducing the rate of people affected by this disease. In order to assist specialists in the diagnosis in bright field smear images, many studies have been developed for the automatic Mycobacterium tuberculosis detection, the causative agent of Tb. To contribute to this theme, a method to bacilli detection associating convolutional neural network (CNN) and a mosaic-image approach was implemented. The propose was evaluated using a robust image dataset validated by three specialists. Three CNN architectures and 3 optimization methods in each architecture were evaluated. The deeper architecture presented better results, reaching accuracies values above 99%. Other metrics like precision, sensitivity, specificity and F1-score were also used to assess the CNN models performance.

Using Convolutional Neural Networks for Classification of Bifurcation Regions in IVOCT Images.

Optical Coherence Tomography (OCT) technology enabled the experts to analyze coronary lesions from high-resolution intravascular images. Studies have shown the relationship between bifurcation regions and a higher occurrence of wall thickening and lesions in these areas. Some level of automation could benefit experts, since examining pullback frames is a laborious and time-consuming task. Although Convolutional Neural Networks (CNN) have shown promising results in classification tasks of medical images, we did not identify the use of CNN's in IVOCT images to classify bifurcation regions in the literature. In this work, we evaluated a CNN architecture in the bifurcation classification task trained with IVOCT images from 9 pullbacks from 9 different patients. We used data augmentation to balance the dataset, due to the low amount of bifurcation-labeled frames. Our classification results are comparable to other works in the literature, presenting better result in AUC (99.70%).

Lumen Segmentation in Optical Coherence Tomography Images using Convolutional Neural Network.

Lumen segmentation in Optical Coherence Tomography (OCT) images is a very important step to analyze points of interest that may help on atherosclerosis diagnostic and treatment. Past studies use many different methods to segment the lumen in IVOCT images, like level set, morphological reconstruction, Markov random fields, and Otsu binarization. Despite Convolutional Neural Networks (CNN) have shown promising results in the image processing area, we did not identify, in the literature, works applying CNN in IVOCT images. In this paper, we present the lumen segmentation using CNN. We evaluated three different CNN architectures. The CNNs were evaluated using three versions from the image dataset, differing from each other by image size (768x768 pixels and 192x192 pixels), and by coordinate system representation (Cartesian and polar). The best results, Accuracy, Dice index and Jaccard index of over 99%, 98% and 97%, respectively, were obtained with the smallest size images represented by polar coordinate system.

A new method for recognizing hand configurations of Brazilian gesture language.

This paper describes a new method for recognizing hand configurations of the Brazilian Gesture Language - LIBRAS - using depth maps obtained with a Kinect® camera. The proposed method comprised three phases: hand segmentation, feature extraction, and classification. The segmentation phase is independent from the background and depends only on pixel depth information. Using geometric operations and numerical normalization, the feature extraction process was done independent from rotation and translation. The features are extracted employing two techniques: (2D)2LDA and (2D)2PCA. The classification is made with a novelty classifier. A robust database was constructed for classifier evaluation, with 12,200 images of LIBRAS and 200 gestures of each hand configuration. The best accuracy obtained was 95.41%, which was greater than previous values obtained in the literature.

Breast tumor classification in ultrasound images using neural networks with improved generalization methods.

Mammography, scintimammography and ultrasound images have been used to increase the specificity of breast cancer image diagnosis. Concerning breast cancer image diagnosis with ultrasound, some results found in the literature show better performance of morphological features in breast cancer lesion differentiation and that a reduced set of features shows a better performance than a large set of features. In this study we evaluated the performance of neural network classifiers, with different training stop criteria: mean square error, early stop and regularization. The last two criteria were developed to improve neural network generalization. Different sets of morphological features were used as neural network inputs. Training sets comprised of 22, 8, 7, 6, 5 and 4 features were employed. To select reduced sets of features, a scalar selection technique with correlation was used. The best results obtained for accuracy and area under the ROC curve were 96.98% and 0.98, respectively. The performance obtained with all 22 features is slightly better than the one obtained with a reduced set of features.

A sputum smear microscopy image database for automatic bacilli detection in conventional microscopy.

In this work, we present an image database for automatic bacilli detection in sputum smear microscopy. The database comprises two parts. The first one, called the autofocus database, contains 1200 images with resolution of 2816 × 2112 pixels. This database was obtained from 12 slides, with 10 fields per slide. Each stack is composed of 10 images, with the fifth image in focus. The second one, called the segmentation and classification database, contains 120 images with resolution of 2816×2112 pixels. This database was obtained from 12 slices, with 10 fields per slice. In both databases, the images were acquired from fields of slides stained with the standard Kinyoun method. In both databases, accordingly to the background content, the images were classified as belonging to high background content or low background content. In all 120 images of segmentation and classification database, the identified objects were enclosed within a geometric shape by a trained technician. A true bacillus was enclosed in a circle. An agglomerated bacillus was enclosed by a rectangle and a doubtful bacillus (the image focus or geometry does not allow a clear identification of the object) was enclosed by a polygon. These marked objects could be used as a gold standard to calculate the accuracy, sensitivity and specificity of bacilli recognition.