Aplicações de modelos computacionais de análise de dados biomédicos em plataformas de dispositivos móveis

Abstract

This work develops CNN­Pulmona and CNN­Cardio, implementations of computational models for optimization of analysis of biomedical data from chest radiography (CXR) and electrocardiogram (ECG), respectively, deployable in mobile device platforms, in addition to a comparison of several methods of signal compression. CNN­Pulmona is an approach for classifying chest X­ray images into three classes: bacterial pneumo nia, viral pneumonia (Covid­19 or other type) and healthy lung. Convolutional Neural Networks are used, based on pre­trained networks in conjunction with a quantization process, by means of the TensorFlow Lite platform method, thereby reducing the com putational cost. The cascade classification method is used, which makes it possible to divide the classifications into different stages; thus, it was possible to obtain 99.16% ac curacy in the classification of images with suspicion of Covid­19. The resulting mobile application program also features a simple and intuitive user interface. In CNN­Cardio, a new method to classify electrocardiogram signals on mobile devices is proposed, which can classify different arrhythmias according to the EC57 standard of the Asso ciation for the Advancement of Medical Instrumentation. A convolutional neural net work was built, trained and validated with the MIT­BIH arrhythmia dataset, in which this database has 5 different classes: normal beat, premature supraventricular beat, pre mature ventricular contraction, ventricular beat fusion, normal and unclassifiable beat. After being trained and validated, the model is submitted to a post­training quantization stage using the TensorFlow Lite conversion method.The results obtained were very sat isfactory, before and after quantization; the convolutional neural network obtained an accuracy of 99%. With the quantization technique, it was possible to obtain a significant reduction in the size of the model, thus enabling the development of the mobile applica tion; this reduction was approximately 90% in relation to the size of the original model. Additionally, the behavior of different signals was compared, when applied to different compression techniques, in order to test and find the best compression techniques for distinct types of biomedical signals, also proving that different types of biomedical sig nals behave distinctly in different types compression of biomedical signals, the results of this comparison of signal compression methods were very satisfactory, demonstrating that different types of compression can be used on signals for better results.