A New Generative Generalised Zero‐Shot Recognition for Chest X‐Ray Images

ABSTRACT

Existing zero‐shot learning (ZSL) models for chest X‐ray images have two main limitations: a lack of annotated training data and ambiguous boundaries in pathological image regions. To address these issues, we propose a locally enhanced generative framework for generalised zero‐shot recognition (GZSR) of chest X‐ray images. First, a diffusion generative model (diffusion‐GAN) is introduced into the GZSR framework and a dual‐discriminator collaborative training strategy is designed to improve the quality of the generated images. Second, to address the issue of blurred pathological regions in chest X‐rays, we have developed a pathological region enhancement module to enhance the visual features of these regions and thereby improve recognition accuracy. Finally, experiments were conducted on two public chest X‐ray datasets: NIH Chest X‐ray 14 and CheXpert. Generalisation performance is specifically validated on the testing classes of the CheXpert datasets. The results of these experiments demonstrate that the proposed model can accurately recognise both seen and unseen classes simultaneously.