Fine-Grained Self-Supervised Learning with Jigsaw puzzles for medical image classification.

Area of Science:

• Medical Imaging Analysis
• Computer Vision
• Machine Learning

Background:

• Classifying fine-grained lesions in medical images presents challenges due to subtle visual differences.
• Training deep neural networks to learn features from such subtle differences is difficult.

Purpose of the Study:

• To introduce a novel Fine-Grained Self-Supervised Learning (FG-SSL) method for improved classification of subtle lesions in medical images.
• To enhance the learning of subtle differences in medical image datasets.

Main Methods:

• The FG-SSL method employs a hierarchical block structure for progressive learning.
• It enforces cross-correlation between fine-grained Jigsaw puzzles and regularized original images to approximate an identity matrix.
• Hierarchical blocks are also applied to supervised learning for enhanced discovery of subtle differences.

Main Results:

• The proposed FG-SSL method demonstrates favorable performance compared to state-of-the-art approaches.
• Experiments were conducted on comprehensive medical image recognition datasets, including ISIC2018, APTOS2019, and ISIC2017.
• The method does not require asymmetric models or negative sampling and is batch size insensitive.

Conclusions:

• The FG-SSL method effectively addresses the challenge of fine-grained lesion classification in medical imaging.
• The progressive learning approach using hierarchical blocks significantly improves the identification of subtle differences.
• FG-SSL offers a robust and efficient solution for medical image analysis tasks.