Description
Background RPE65-biallelic mutation-associated inherited retinal disease (RPE65‑IRD) presents with distinct retinal imaging features that evolve with disease progression, including whitish dots (WD), marble‑like pathology (ML), hypopigmented areas (HA), patchy chorioretinal atrophy (PCA), pigmented spicules (PS), and spotty irregular hyperpigmentation (SIHP). Quantification of these features may provide objective biomarkers for disease monitoring.
Aim To develop and evaluate deep learning models for segmentation of six RPE65‑IRD-associated imaging features.
Method A retrospective cohort of 23 patients (46 eyes) with genetically confirmed RPE65 variants was studied. A total of 178 ultra‑widefield (133°) true‑colour fundus images were graded by a clinician. A feature-specific curation scheme was implemented. Deep learning models were trained using full‑image and patch‑based approaches on the complete and curated datasets with five‑fold cross‑validation. Performance was evaluated using the Sørensen-Dice coefficient, sensitivity, and specificity, with comparisons assessed using the Wilcoxon signed‑rank test. Qualitative validation used the DINOv3 vision foundation model.
Results Patch‑based training significantly improved segmentation of small features (PS, SIHP; p < 0.05). Image‑level classification showed no comparable gains, reflecting sensitivity to annotation noise. DINOv3 analysis confirmed visual similarity between PS and SIHP.
Conclusion Targeted data curation and patch‑based learning mitigate label noise and enable robust segmentation of peripheral RPE65‑IRD biomarkers, supporting objective assessment of disease status.
Lay Abstract
RPE65-related inherited retinal disease is a rare eye condition that affects the retina and gradually worsens over time. Eye images can show different visible changes, such as small white spots, marble-like patterns, and darkened areas. As the disease progresses, these changes may turn into areas of retinal damage and darker pigment patterns. Measuring these changes could help doctors better understand and monitor the condition.
In this study, we developed and tested artificial intelligence (AI) tools to identify and measure six key retinal features linked to this disease. We analysed 178 wide-angle colour retinal images from 23 patients (46 eyes) with genetically confirmed disease. All images were reviewed by a clinician and carefully organised to reduce errors. The AI models were trained using both full images and smaller image sections, and their performance was compared.
Training the AI on smaller image sections improved detection of very small features, while models trained only on full images performed less well. Some errors were related to differences in how consistently humans can label these features. Overall, careful data selection and targeted AI training improved detection of retinal changes and may support more reliable monitoring of disease progression.
| Lay Title | AI Helps Track Retinal Changes in a Rare RPE65-Related Eye Disease |
|---|---|
| Role | Research Assistant |