Automated Nonperfusion Quantification in Diabetic Retinopathy on Ultra-Widefield Swept-Source OCT Angiography

To evaluate the performance of a customized deep learning algorithm for automated segmentation of nonperfusion area (NPA) on ultrawide-field swept-source optical coherence tomography angiography (UWF SS-OCTA) and its utility in diabetic retinopathy (DR) severity assessment. Design Cross-sectional study SubjectsA total of 180 eyes from 122 participants representing all grades of DR severity. MethodsWe developed a convolutional neural network (CNN) based on a multi-scale U-Net backbone with squeeze-and-excitation attention for segmentation of NPAs on en face SS-OCTA all-retinal-layer (ARL) images from three scan patterns: 6 × 6 mm, 12 × 12 mm, and 29 × 24 mm.Ground truth annotations of NPAs and non-gradable area (NGA) on en face OCTA images were generated by two independent graders and adjudicated by a vitreoretinal specialist.A corresponding en face structural OCT was incorporated to distinguish true NPAs from shadow artifacts.Segmentation outputs included NPA, NGA and shadow artifacts.Pixel-level accuracy was assessed with the F1 score.Nonperfusion index (NPI) was defined as NPA/gradable area.The level of agreement between human-labelled and AI predicted NPI was analyzed using Bland-Altman analysis. J o u r n a l P r e -p r o o f Main Outcome MeasuresAlgorithm F1 score and NPI. ResultsThe algorithm for NPA segmentation achieved a mean F1 score of 0.82 ± 0.01 in 6 × 6 mm, 0.84 ± 0.03 in 12 × 12mm and 0.83 ± 0.02 in 29 × 24mm, with no significant difference across fields of view (p = 0.12).Algorithm-derived NPI strongly agreed with expert grading (ICC > 0.99).Both human-and algorithm-derived NPI increased progressively with increased DR severity in all scan patterns demonstrated by the Kruskal-Wallis test.(6 × 6 mm: human: p = 0.02; algorithm: p = 0.03; 12 × 12 mm: algorithm p < 0.001; human p < 0.001; 29 × 24 mm: algorithm：p < 0.001; human：p < 0.001) with the largest magnitude of increase in 29 × 24 mm scans.The algorithm for FAZ segmentation also achieved a mean F1 score of 0.88 ± 0.03 for 6 x 6 mm images and 0.85 ± 0.02 for 12 x 12 mm images. ConclusionThis deep learning algorithm validated on single-scan UWF SS-OCTA for automated NPA segmentation and quantification.It demonstrates high accuracy and scalability across multiple scan sizes, supporting its potential integration into objective DR OCTA biomarker analysis.