Multi-Expert Deep Networks for Multi-Disease Detection in Retinal Fundus Images

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging Analysis

Background:

• Automatic diagnosis of eye diseases from retinal fundus images presents significant challenges.
• Public datasets often feature uneven label distribution, under-representing rare diseases.
• Accurate multi-label classification is crucial for comprehensive eye disease detection.

Purpose of the Study:

• To propose a robust training pipeline for multi-label classification of eye diseases from retinal images, addressing uneven sample distribution and difficulty.
• To enhance the performance of diagnostic models for both common and rare eye conditions.
• To improve the accuracy and reliability of automated eye disease detection systems.

Main Methods:

• Implemented a training pipeline incorporating inverse-frequency class weighting to balance under-represented and over-represented samples.
• Adjusted class weights iteratively based on aggregated loss to focus training on difficult samples.
• Utilized a novel Heuristic Stacking algorithm for ensembling models to improve multi-label predictions.

Main Results:

• Achieved an 88.24% accuracy score on the Retinal Image Analysis for Multi-Disease Detection (RIADD)-2021 dataset.
• Demonstrated competitive performance against top-ranked methods in the challenge.
• Ablation studies confirmed the effectiveness of the Heuristic Stacking ensemble method over classical approaches.

Conclusions:

• The proposed training pipeline effectively addresses challenges in multi-label classification of eye diseases with uneven data distribution and difficulty.
• Heuristic Stacking offers a superior ensemble method for multi-label prediction in medical image analysis.
• The approach shows significant promise for improving automated diagnosis of eye diseases, including rare conditions.