Asynchronous feature regularization and cross-modal distillation for OCT based glaucoma diagnosis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible vision loss, necessitating early diagnosis.
• Interpreting ophthalmic data for glaucoma diagnosis is time-consuming and resource-intensive.
• Automated diagnostic methods are crucial for efficient and accurate glaucoma detection.

Purpose of the Study:

• To develop a novel deep learning framework to enhance glaucoma diagnosis using optical coherence tomography (OCT) data.
• To improve the performance of OCT-based models by transferring knowledge from visual field (VF) testing.
• To create an automated system that assists ophthalmologists in fast and accurate glaucoma diagnosis.

Main Methods:

• A novel deep learning framework was developed to enhance OCT model performance.
• A cross-modal knowledge transfer method was designed, integrating a distillation loss and an asynchronous feature regularization (AFR) module.
• The framework leveraged OCT and VF data for knowledge transfer, then evaluated on an external OCT dataset.

Main Results:

• The proposed method achieved an accuracy of 87.4% and an AUC of 92.3% using only OCT inputs, surpassing previous OCT-only models.
• The performance of the enhanced OCT model was comparable to joint OCT and VF models.
• Evaluation on an external dataset confirmed the model's effectiveness and generalization capability.

Conclusions:

• The developed deep learning framework effectively enhances glaucoma diagnosis accuracy using OCT data.
• Cross-modal knowledge transfer from VF data significantly improves OCT-based glaucoma detection.
• The model demonstrates strong generalization, offering a promising tool for clinical application in early glaucoma diagnosis.