CONICET | Buscador de Institutos y Recursos Humanos

Medical images provide tissue representation assisting medical experts in pathology diagnosis and also help-ing to interpret the study of human anatomy. Chest X-ray images provides a general framework for the detection of anomalies related to possible diseases, becoming essential, for example, for the diagnosis of pneumonia, detection of masses, and, in the last two years, for the detection of medical conditions caused by COVID-19. Neural networks based on deep-learning, in particular, Convolutional Neural Networks (CNN) have gained momentum in recent years, becoming central tools for image classification. Approaches based on transfer-learning, in which previously trained network layers are reused (typically the ones correspond-ing to the feature extraction phase), have made it possible to overcome the limitations of the CNNs related to the size of the training datasets and also the computational cost. In this work, we present a study of transfer-learning approaches on CNNs for the detection of pathologies in tchest X-ray. We used the public database ChestX-ray14, which contains more than 100,000 images with labels related to 14 possible pathol-ogies. We analyse and compare CNNs based on the ImageNet dataset, such as VGG19, RESNET, Incep-tionv3, among others, considering different schemes for the classification phase and training approaches, including data-augmentation. We explore their ability to classify pure pathologies versus no-finding and classification between different pathologies. The results indicate an acceptable performance in most of the cases tested, indicating that transfer-learning-based approaches are a valid initial path for the use of CNNs in cases where there is not enough labelled images.