Harnessing the wisdom of a radiologist: Texture-aware curriculum self-supervised learning for thorax disease classification.

The article focuses on a novel texture-aware self-supervised learning framework for thorax disease classification in medical imaging, specifically chest X-rays. It introduces a curriculum learning strategy that organizes large-scale unlabeled radiographs by texture complexity using the Gray-Level Co-occurrence Matrix (GLCM), progressively training the model from simple to complex texture patterns. The framework integrates a lightweight CNN-based patch embedding module within a Vision Transformer (ViT) architecture and employs a GLCM-based constraint during masked radiograph reconstruction to enhance texture feature extraction. Experimental results on the NIH Chest X-Ray and Stanford CheXpert datasets demonstrate that this approach outperforms existing state-of-the-art methods in thorax disease classification, highlighting the importance of texture information and curriculum learning in medical image analysis.