Automated Identification and Segmentation of Diabetic Macular Edema Subtypes Using Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Diabetic macular edema (DME) is a severe complication of DR.
• Accurate classification and segmentation of DME subtypes are crucial for effective treatment.

Purpose of the Study:

• To develop a deep learning algorithm for classifying DME subtypes.
• To segment lesions associated with DME using structural optical coherence tomography (OCT) images.
• To evaluate the algorithm's performance and cross-device generalizability.

Main Methods:

• A dataset (DME-Seg) of 3120 DME OCT B-scan images was curated from 823 eyes across four OCT devices.
• Expert ophthalmologists provided multi-label annotations, including DME subtype classification and pixel-level segmentation masks.
• The YOLO11x-Seg model was fine-tuned on the DME-Seg dataset for detection and segmentation tasks.

Main Results:

• The fine-tuned model achieved high performance in lesion detection (mAP50(B) = 0.82) and segmentation (mAP50(M) = 0.84).
• Dice coefficients for detection and segmentation were 0.82 ± 0.20 and 0.79 ± 0.18, respectively.
• The model demonstrated promising cross-device generalizability.

Conclusions:

• The developed deep learning algorithm and DME-Seg dataset show potential for automated DME subtype detection and segmentation.
• This AI-assisted tool can enhance the precision of DME diagnosis and treatment planning.
• The findings support the clinical utility of AI in ophthalmology for managing diabetic eye disease.