Quickly diagnosing Bietti crystalline dystrophy with deep learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Bietti crystalline dystrophy (BCD) is a challenging autosomal recessive inherited retinal disease (IRD).
• Early and precise diagnosis of BCD is crucial for patient management.
• Current diagnostic methods may not be sufficient for timely detection.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for diagnosing BCD and classifying its clinical stages.
• To utilize ultra-wide-field (UWF) color fundus photographs (CFPs) for BCD detection.
• To establish an automated diagnostic and grading system for BCD in a Chinese population.

Main Methods:

• Development of DL models (ResNeXt, Wide ResNet, ResNeSt) using UWF-CFPs.
• Classification of images into BCD, retinitis pigmentosa (RP), or normal categories.
• Grading of BCD patients into three clinical stages.
• Evaluation of model performance using accuracy and confusion matrices.
• Verification of diagnostic interpretability through heatmaps.

Main Results:

• DL models achieved good classification performance for BCD detection and staging.
• The study established the largest BCD database for the Chinese population.
• Demonstrated the potential efficacy of an automated DL algorithm for BCD diagnosis and grading.

Conclusions:

• Deep learning models can effectively diagnose Bietti crystalline dystrophy using UWF fundus photography.
• An automated DL approach shows promise for quick diagnosis and clinical staging of BCD.
• This study provides a valuable resource and methodology for BCD research in the Chinese population.