Aprendizagem Autossupervisionada Aplicada à Detecção de Infecção Pulmonar Causada pela COVID-19 em Radiografias de Tórax

Abstract

The state-of-the-art of automatic image classification comprises convolutional neural networks as well as other deep learning techniques. However, for supervised training of deep neural networks, in general, a large number of annotated images is necessary. In the case of new diseases, as occurred for COVID-19 at the beginning of the pandemic, the necessary volume of annotated images to train deep learning algorithms can remain scarce for a long time. In a scenario of the scarcity of labeled data for the training supervised deep algorithms, a potential alternative comprises the use of self-supervised deep learning methods. The main advantage of self-supervised approaches is that they dispense labels. Several of these methods have emerged in recent years and have been shown to be efficient to produce an embedding space where similar objects are close to each other and dissimilar objects are distant to each other. This feature allows embedding spaces to be used as input to simple classifiers such as linear classifiers. Among the advantages obtained in this approach, it can be mentioned that linear classifiers can be trained in a supervised way from a set of labeled data that is much scarcer than what would be necessary to train, from scratch, a supervised classifier based on deep learning. However, in the literature, research involving applications and comparative studies on the use of self-supervised methods for medical imagery in the context of COVID-19 is rare. The emergence of new diseases added to the consequent scarcity of medical images duly annotated in the face of an unknown disease scenario corroborates the potential benefit of computational techniques that use self-supervised methods. Thus, this work evaluates the performance of the following selfsupervised learning methods in the context of chest radiography classification: MoCo, SimSiam, BYOL, and SimCLR. As a dataset, COVID-X was used, which contains cases of pneumonia, and pneumonia due to Covid-19, in addition to images showing healthy patients. The area under the ROC curve. The area under the ROC curve achieved by the MoCo, SimCLR, BYOL and SimSiam frameworks was 0.9757, 0.8903, 0.9765, and 0.9723, respectively. The overall accuracy for these methods, respectively, was 0.8956, 0.8865, 0.8962, and 0.8357.