MTCLF: A multitask curriculum learning framework for unbiased glaucoma screenings.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma causes irreversible optic nerve damage, necessitating accurate automatic screening.
• Current algorithms face limitations due to imbalanced datasets, hard samples, and lack of clinical interpretability.

Purpose of the Study:

• To develop an unbiased glaucoma screening algorithm with enhanced interpretability.
• To address limitations of existing methods in handling complex fundus diseases and imbalanced data.

Main Methods:

• Proposed a multitask curriculum learning framework (MTCLF) using a teacher-student network structure.
• Implemented a dual-branch CNN with a collaborative learning module for simultaneous diagnosis and evidence map prediction.
• Utilized curriculum coefficients to guide training, balance sample contribution, reduce bias, and mine hard samples.

Main Results:

• MTCLF achieved superior performance on the LAG dataset with accuracy (0.967), sensitivity (0.961), specificity (0.970), AUC (0.996), and F2-score (0.958).
• Results surpassed those of existing state-of-the-art methods in glaucoma diagnosis.

Conclusions:

• MTCLF demonstrates optimal performance for unbiased glaucoma diagnosis.
• The generated evidence maps aid clinicians in identifying subtle lesion areas, improving clinical application potential.