Coarse-to-fine visual representation learning for medical images via class activation maps

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep neural networks require large, meticulously annotated datasets, which are costly and time-consuming to acquire in the medical field.
• Coarse labels, such as binary normal/abnormal classifications, offer a cost-effective alternative for training medical image analysis models.
• Developing methods to leverage these coarse labels for learning robust and transferable representations is crucial for advancing medical AI.

Purpose of the Study:

• To investigate the utility of coarsely labeled datasets for learning transferable representations in medical imaging.
• To propose and evaluate CAMContrast, a novel two-stage representation learning framework designed to utilize binary labels effectively.
• To demonstrate the generalizability and efficiency of learned representations across various downstream medical imaging tasks.

Main Methods:

• CAMContrast employs a two-stage representation learning approach for medical images.
• It utilizes class activation maps (CAMs) generated from binary labels (normal vs. abnormal) to create positive views for contrastive learning.
• The framework optimizes a learning objective to maximize agreement within image-heatmap pairs, learning fine-grained, generalizable representations.

Main Results:

• Empirical validation on public datasets (fundus photographs, chest X-rays) for classification and segmentation tasks.
• CAMContrast demonstrated superior performance compared to other self-supervised and supervised pre-training methods.
• The method showed significant advantages in both data efficiency and downstream task performance.

Conclusions:

• Coarsely labeled datasets are valuable for learning transferable representations in medical imaging.
• CAMContrast effectively leverages binary labels and CAMs to learn high-quality, generalizable medical image representations.
• The proposed framework offers a promising direction for efficient and effective medical AI development, especially in low-label scenarios.