Domain adaptation applied to the classification of Covid-19 chest CT images.

Abstract

One of the most common uses of convolutional neural networks is image classification. This can be performed in a supervised, semi-supervised and unsupervised manner depending on the available data. This task is specially important in the field of Computer Aided Diagnosis, where many biomedical applications of deep learning are developed and used to aid medical professionals. As the Covid-19 progressed across the globe, multiple image classification models were developed, most of them using well known networks such as ResNet and DenseNet as a backbone, and in order to save time, were pre-trained on the ImageNet dataset and then fine-tuned to the Covid-19 disease images. We used a public Covid-19 dataset composed of 7305 subjects in a pre-training and finetuning based training framework to classify the disease and GradCAM to visualize saliency maps and visually evaluate if the pipelines in the framework use medically relevant information in the CT slices to make the classification choice. While the models perform along the line of the state of the art in terms of accuracy, with over 0.9 accuracy, the visualization of saliency maps show that they do not necessarily find relevant imaging findings in the images. This shows the need for more transparency when showing classification results in tasks related to the medical imagining domain, in order to build trust with medical professionals and advance in the computer aided diagnosis field. Other classification methods are proposed, such as adversarial learning and attention based learning, in order to overcome the shortcut learning problem to improve the saliency maps, such as adversarial and self supervised learning.