Efficient labeling of retinal fundus photographs using deep active learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) classification relies on deep learning models.
• Labeling large datasets for DR classification is time-consuming and expensive.
• Optimizing label efficiency is crucial for developing robust DR detection systems.

Purpose of the Study:

• To compare the performance of four deep active learning (DAL) approaches.
• To optimize label efficiency in training deep learning models for diabetic retinopathy (DR) classification.

Main Methods:

• Utilized 88,702 retinal fundus images from 44,351 patients (EyePACS dataset).
• Compared four DAL methods (entropy sampling, BALD, core set, adversarial) against random sampling.
• Evaluated models using Cohen's kappa (CK) and AUC on internal and independent test sets.

Main Results:

• Three of four DAL methods significantly improved multiclass DR classification (CK differences: 0.051-0.053).
• Improvements generalized to an independent test set (CK differences: 0.126-0.135).
• No significant improvements were observed for binary classification; ADV performed similarly to random sampling.

Conclusions:

• Uncertainty-based and feature descriptor-based DAL methods outperform random sampling for multiclass DR classification.
• DAL enhances label efficiency for training DR classification models.
• Binary classification performance did not benefit from DAL methods in this study.