Few-shot diagnosis of chest x-ray images using auxiliary information guided semi-deterministic infinite mixture prototypes

Abstract

We propose a few-shot learning (FSL) approach for the diagnosis of chest x-ray images. Our model can be trained with a small number of annotated data by utilizing auxiliary semantic information about the abnormalities under consideration. In our design, we consider the fact that because of various factors, there may be variations in the visual characteristics of an abnormality in x-rays. Hence, in a multi-label dataset, it is challenging to represent data points with a particular abnormality in one cluster based on visual features. Our few-shot learning approach dynamically generates multiple clusters to accurately represent a particular abnormality. The generation of multiple clusters is achieved using a semi-deterministic infinite mixture prototype method. The clustering process is guided by semantic information corresponding to the abnormalities. Thus, our method aims to create a discriminative representation for x-ray images utilizing semantic information about the abnormalities under consideration. Experiments on publicly available chest x-ray datasets show the efficacy of the proposed method for the diagnosis of chest x-ray images. Our code is publicly available in this repository.²