Diffusion model with central annotation propagation for segmentation of ellipsoid zone maps on OCT scans of retinitis pigmentosa.

We present a fully automatic algorithm to segment the ellipsoid zone (EZ) on optical coherence tomography (OCT) images of retinitis pigmentosa (RP). We introduce Diffusion Projection U-Net (DPUN), a method to generate en face EZ maps from 3D OCT scans by leveraging diffusion models' state-of-the-art generative capabilities. We further propose Central Annotation Propagation (CAP), a semi-automatic approach that incorporates the segmentation of the central scan to segment all scans in a volume using a memory-reading paradigm. We demonstrated the effectiveness of our methods on a dataset consisting of 60 eyes from 60 participants enrolled in an international, multicenter, phase 3 RP clinical trial (NCT05537220) imaged with spectral-domain OCT at the baseline visit. DPUN outperformed all comparison methods for fully automatic segmentation. DPUN-CAP, which utilized central annotations, achieved the best performance of all models evaluated. The proposed methods produced measurements that were consistent with manual measurements of the EZ area, highlighting their potential utility for clinical trials in which EZ area is a key anatomic endpoint and for monitoring EZ loss progression and treatment response in clinical practice.