Hierarchical-Instance Contrastive Learning for Minority Detection on Imbalanced Medical Datasets

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models face challenges with imbalanced datasets and limited data, particularly in medical imaging where rare diseases represent minority classes with diverse distributions.
• Insufficient labels and unseen cases further complicate training for minority classes in medical datasets.

Purpose of the Study:

• To propose a novel Hierarchical-instance Contrastive Learning (HCLe) method for effective minority class detection in medical imaging.
• To address data imbalance and data-hungry issues by training solely on majority class data.

Main Methods:

• HCLe employs a two-branch architecture: an auto-encoder with constraint functions for image-level feature extraction and a contrastive learning network for hierarchical sample feature consistency.
• A diverse mini-batch strategy refines the method for identifying minority classes under various conditions.

Main Results:

• The proposed HCLe method was evaluated on three diverse medical datasets (different diseases and modalities).
• Experimental results demonstrated that HCLe significantly outperforms current state-of-the-art methods in minority class detection.

Conclusions:

• HCLe offers a robust solution for minority detection in medical imaging, effectively overcoming common deep learning limitations.
• The method shows promise for improving the diagnosis of rare diseases and other minority classes in medical datasets.