GraphXCOVID: Explainable Deep Graph Diffusion Pseudo-Labelling for\n Identifying COVID-19 on Chest X-rays

Abstract

Can one learn to diagnose COVID-19 under extreme minimal supervision? Since\nthe outbreak of the novel COVID-19 there has been a rush for developing\nArtificial Intelligence techniques for expert-level disease identification on\nChest X-ray data. In particular, the use of deep supervised learning has become\nthe go-to paradigm. However, the performance of such models is heavily\ndependent on the availability of a large and representative labelled dataset.\nThe creation of which is a heavily expensive and time consuming task, and\nespecially imposes a great challenge for a novel disease. Semi-supervised\nlearning has shown the ability to match the incredible performance of\nsupervised models whilst requiring a small fraction of the labelled examples.\nThis makes the semi-supervised paradigm an attractive option for identifying\nCOVID-19. In this work, we introduce a graph based deep semi-supervised\nframework for classifying COVID-19 from chest X-rays. Our framework introduces\nan optimisation model for graph diffusion that reinforces the natural relation\namong the tiny labelled set and the vast unlabelled data. We then connect the\ndiffusion prediction output as pseudo-labels that are used in an iterative\nscheme in a deep net. We demonstrate, through our experiments, that our model\nis able to outperform the current leading supervised model with a tiny fraction\nof the labelled examples. Finally, we provide attention maps to accommodate the\nradiologist's mental model, better fitting their perceptual and cognitive\nabilities. These visualisation aims to assist the radiologist in judging\nwhether the diagnostic is correct or not, and in consequence to accelerate the\ndecision.\n

Ähnliche Arbeiten