A deep network using coarse clinical prior for myopic maculopathy grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Pathological myopia (PM) is a leading cause of blindness worldwide.
• Myopic maculopathy grading is crucial for assessing PM severity and progression.
• Timely treatment based on accurate grading can prevent myopia-induced blindness.

Purpose of the Study:

• To develop and evaluate a novel feature fusion framework for grading myopic maculopathy.
• To leverage fundus image priors for improved diagnostic accuracy in pathological myopia.
• To enhance the prediction of severity and progression for effective PM management.

Main Methods:

• A feature fusion framework integrating tessellated fundus and brightest region priors was proposed.
• Prior knowledge extraction module used traditional image processing for lesion localization.
• Feature fusion module incorporated priors as global and local constraints within a deep learning network.
• Rank loss was designed to improve classification score continuity.

Main Results:

• The framework achieved 0.8921 five-grade accuracy on a private dataset from Beijing TongRen Hospital (714 images).
• On the Pathological Myopia (PALM) dataset, the framework achieved an AUC of 0.9981 for two-class grading.
• The proposed method demonstrated strong performance in myopic maculopathy grading.

Conclusions:

• The feature fusion framework shows significant potential for accurate myopic maculopathy grading.
• Integrating image priors enhances deep learning models for pathological myopia assessment.
• This approach can aid clinicians in timely treatment decisions to prevent blindness.