Long-tailed multi-label retinal disease classification using alternate group training and gradient-based...

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Ocular diseases are a leading cause of vision impairment, requiring accurate and timely diagnosis.
• Current computer-aided diagnostic methods struggle with class imbalance and label co-occurrence in ocular disease data.
• The long-tailed distribution of ocular diseases complicates accurate classification and diagnosis.

Purpose of the Study:

• To propose a novel alternate group training strategy to address the multi-label long-tailed data distribution problem in ocular disease diagnosis.
• To reduce the challenges posed by class imbalance and label co-occurrence in medical image analysis.
• To enhance the performance of computer-aided diagnostic systems for ocular conditions.

Main Methods:

• Data partitioning into semantic groups to mitigate class imbalance and label co-occurrence.
• Alternate training of a teacher network using a gradient-based self-weighted loss.
• Training a student model on the original dataset guided by the teacher network with a weighted class-balanced distillation loss.

Main Results:

• The proposed alternate group training strategy significantly outperforms existing methods on a public ocular disease dataset.
• The method effectively alleviates class-wise imbalance and instance-wise label co-occurrence.
• Performance improvements were also observed when extending the single-teacher model to a multi-teacher framework.

Conclusions:

• The novel alternate group training strategy is a highly effective approach for tackling multi-label, long-tailed data distribution problems in medical diagnostics.
• This method offers a promising solution for improving the accuracy and robustness of computer-aided ocular disease detection.
• The strategy's adaptability to multi-teacher models suggests broad applicability in complex diagnostic scenarios.