Integrating Self-Supervised Learning with Nonlinear Classifiers in Lightweight Swin Transformer for X-Ray Image Classification

Abstract

In this paper, we present a new approach about the integration Self-Supervised Learning with nonlinear Clas-sifiers in Lightweight Swin Transformer (SSLnC - LSwinT)  for  improving X-ray image classification performance. Our approach leverages unlabeled data to address the issue of labeled data scarcity in the medical field by using self-supervised learning (SSL) to extract features. One of our key contributions is the introduction of the Lightweight SwinT architecture, a more compact variant of SwinT, designed to enhance computational efficiency, reduce model complexity, and shorten training time. To further improve classification efficiency, we propose the integration of a nonlinear classifier instead of a traditional linear classifier in Lightweight SwinT. Experimental results underscore the impact of our contribu-tions, demonstrating significant reductions in model training time and notable improvements in classification performance. Our proposed method, which integrates SSL based on LSwinT with a nonlinear LightGBM classifier, achieves an accuracy of up to 87%, improving by 1.8% over the non-LightGBM SwinT version and cutting training time significantly (3:23:00 vs. 7:37:29) compared to the original SwinT architecture.