Hierarchical discriminative learning improves visual representations of biomedical microscopy.

Area of Science:

• Biomedical imaging
• Computer vision
• Machine learning

Background:

• Self-supervised representation learning (SSL) is crucial for computer vision in medicine.
• Existing SSL methods applied to whole-slide images (WSIs) overlook the patient-slide-patch hierarchy and require strong data augmentations.

Purpose of the Study:

• To develop a data-driven method (HiDisc) that leverages the hierarchical structure of biomedical data for improved self-supervised visual representation learning.
• To address limitations of current SSL methods in handling the inherent data hierarchy and reducing reliance on aggressive data augmentations.

Main Methods:

• HiDisc utilizes a self-supervised contrastive learning framework.
• It defines positive patch pairs based on common ancestry within the patient-slide-patch hierarchy.
• A unified objective learns discriminative features across patch, slide, and patient levels.

Main Results:

• HiDisc pretraining significantly outperforms state-of-the-art SSL methods in cancer diagnosis and genetic mutation prediction tasks.
• The method learns high-quality visual representations by utilizing natural patch diversity without requiring strong data augmentations.

Conclusions:

• HiDisc effectively learns robust visual representations by respecting the hierarchical nature of biomedical microscopy data.
• This approach advances self-supervised learning for clinical applications, offering improved performance and reduced data augmentation needs.