Automated FAZ segmentation and diabetic retinopathy classification using OCTA images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) diagnosis relies on identifying alterations in the foveal avascular zone (FAZ), a key biomarker.
• Accurate FAZ segmentation in optical coherence tomography angiography (OCTA) images is crucial for DR assessment.

Purpose of the Study:

• To develop and evaluate an automated deep learning framework for FAZ segmentation and DR classification using OCTA images.
• To explore the feasibility of using deep learning for early and accurate DR diagnosis.

Main Methods:

• A two-step deep learning pipeline was employed, integrating DeepLabv3+, EfficientNetB0, SE blocks, and ASPP for FAZ segmentation.
• A GoogLeNet-based CNN was utilized for classifying DR stages (normal, NPDR, PDR) based on segmented FAZ images.
• Data augmentation and SMOTE were applied to enhance classification performance, with 5-fold cross-validation used.

Main Results:

• The FAZ segmentation network achieved a high Dice Similarity Coefficient (DSC) of 97.5%.
• The classification model demonstrated 100% AUC for binary (normal vs. DR) and 87% AUC for three-class (normal, NPDR, PDR) classification.
• The dataset included 253 OCTA scans from 161 participants with varying stages of DR.

Conclusions:

• The developed automated framework shows significant potential as an assistive tool for clinicians.
• This system can enable earlier and more accurate diagnosis of diabetic retinopathy from OCTA imaging.
• Integration into clinical workflows could improve patient outcomes through timely DR detection.