Purpose To develop and validate OCT angiography (OCTA) Artifact Removal Model (OCTA-ARM), a diffusion-based deep learning framework for simultaneously removing motion, shadow, and defocus artifacts on OCTA images. Design Retrospective development and validation study. Subjects A total of 21 389 OCTA superficial capillary plexus images from 2475 eyes of 1266 patients were included. The paired evaluation dataset comprised 279 matched artifact-affected and artifact-free image pairs. An additional 3260 images from 574 eyes of 376 patients with diabetes were analyzed for correlation with diabetic retinopathy (DR) severity. Methods OCT angiography ARM employs a 2-stage training pipeline utilizing CycleGAN-based generators to synthesize shadow and defocus artifacts, and randomized row-wise pixel shifts for motion artifacts. The Brownian bridge diffusion model architecture was used for artifact removal. Model performance was evaluated using image quality metrics (structural similarity index measure, peak signal-to-noise ratio, Fréchet inception distance, and learned perceptual image patch similarity). Agreement of quantitative OCTA metrics between artifact-affected and artifact-free image pairs was assessed using intraclass correlation coefficients (ICCs) before and after OCTA-ARM processing. Spearman correlation coefficients were calculated to evaluate correlations between OCTA metrics and International Clinical Diabetic Retinopathy (ICDR) grades. Main Outcome Measures Image quality metrics, ICC for quantitative OCTA metrics, and correlation with DR severity. Results After OCTA-ARM reconstruction, all 200 “gradable but unmeasurable” images were upgraded to “measurable,” and 66 of 79 (83.5%) “ungradable” images were upgraded to “measurable” status. Intraclass correlation coefficient values between artifact-free and artifact-affected image pairs for quantitative metrics significantly improved after artifact removal using OCTA-ARM (all P < 0.0001): foveal avascular zone (FAZ) area (from 0.771 to 0.909), vessel density (VD; from 0.416 to 0.673), fractal dimension (FD; from 0.385 to 0.757), and skeleton density (SD: from 0.372 to 0.757). Correlations between OCTA metrics and ICDR grades consistently strengthened after reconstruction (all P < 0.001): FAZ area (r: 0.116 to 0.203), nonperfusion area (r: 0.253 to 0.346), VD (r: –0.258 to –0.355), FD (r: –0.337 to –0.472), and SD (r: –0.327 to –0.459). Conclusions OCT angiography ARM effectively removes multiple artifact types from OCTA images, significantly improving image quality, enhancing the reliability of quantitative metrics, and strengthening clinical correlations with DR severity. Financial Disclosure(s) Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
