TNCB: Tri-Net With Cross-Balanced Pseudo Supervision for Class Imbalanced Medical Image Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are crucial for medical image analysis but struggle with limited labeled data and class imbalance.
• Existing semi-supervised learning (SSL) methods often fail to address class imbalance, leading to prediction bias towards majority classes.
• Training bias also arises from suboptimal strategies in pseudo-label generation and utilization within current SSL frameworks.

Purpose of the Study:

• To introduce a novel semi-supervised learning (SSL) framework, Tri-Net with Cross-Balanced pseudo supervision (TNCB), designed to overcome label scarcity and class imbalance in medical image classification.
• To reduce prediction bias by enabling the teacher model to focus on minority classes.
• To mitigate training bias through an adaptive cross-loss function for improved knowledge extraction from unlabeled data.

Main Methods:

• Proposed the Tri-Net with Cross-Balanced pseudo supervision (TNCB) framework, comprising two student networks and a teacher network with an adaptive balancer.
• Implemented a virtual optimization strategy to enhance the teacher model's robustness against class imbalance.
• Utilized cross-balanced pseudo supervision with an adaptive cross-loss function to minimize training bias and leverage unlabeled data.

Main Results:

• TNCB demonstrated superior performance compared to state-of-the-art SSL methods across four diverse datasets.
• The framework effectively addressed challenges related to different diseases, image modalities, and varying imbalance ratios.
• Consistent improvements in classification accuracy highlight the effectiveness of TNCB.

Conclusions:

• TNCB is an effective and robust SSL framework for imbalanced medical image classification.
• The proposed methods successfully reduce both prediction and training bias inherent in existing SSL approaches.
• TNCB offers a promising solution for enhancing deep learning applications in clinical settings with limited and imbalanced medical data.