Hierarchical Knowledge Guided Learning for Real-World Retinal Disease Recognition.

Area of Science:

• Ophthalmology AI
• Medical Image Analysis
• Deep Learning

Background:

• Medical datasets, particularly in ophthalmology, often have long-tailed distributions, with rare diseases underrepresented.
• Deep learning models struggle to generalize to rare diseases due to limited training samples.
• Retinal image analysis faces challenges with multi-label classification (co-occurring diseases) and imbalanced data.

Purpose of the Study:

• To develop a novel deep learning method for accurate retinal disease recognition from long-tailed fundus image datasets.
• To address challenges of rare disease identification and label co-occurrence in medical AI.
• To improve the generalization ability of deep neural networks on imbalanced ophthalmology data.

Main Methods:

• Utilized hierarchy-aware pre-training leveraging ophthalmology prior knowledge to enhance feature representation.
• Implemented an instance-wise class-balanced sampling strategy to manage label co-occurrence in long-tailed datasets.
• Introduced a hybrid knowledge distillation technique for training less biased representations and classifiers.

Main Results:

• The proposed method demonstrated superior performance on four diverse datasets (over one million fundus images).
• Achieved state-of-the-art recognition accuracy, significantly outperforming existing methods.
• Showed particular effectiveness in recognizing rare retinal diseases.

Conclusions:

• The novel deep learning approach effectively handles long-tailed distributions and label co-occurrence in retinal fundus image analysis.
• The method offers a significant advancement for AI in ophthalmology, improving diagnosis of both common and rare diseases.
• This work provides a robust solution for training deep neural networks on imbalanced medical data, enhancing diagnostic capabilities.