Deep Learning for Diagnosis of Choroideremia and USH2A-Associated Rod-Cone Dystrophy Using Macular OCT Volumes

To develop and validate deep learning models to differentiate choroideremia (CHM), USH2A-associated rod-cone dystrophy, and healthy controls using macular OCT volumes.

Retrospective diagnostic study with external validation.

Two hundred sixty-two participants contributed 535 macular OCT volumes acquired on a Heidelberg Spectralis (10 165 B-scans): 99 participants had CHM, 96 had USH2A-associated disease, and 67 were healthy controls.

Volumes acquired between 2012 and 2021 were standardized to 19 B-scans and split at the patient level into training (76%), validation (12%), and test (12%) sets. Three strategies were compared: (1) a slice-based convolutional neural network (CNN) (ResNet-50) with majority-vote aggregation; (2) a Mixture-of-Experts (MoE) model with learned slice weighting; and (3) a hybrid CNN-Transformer integrating per-slice features across the volume. Two ophthalmologists masked to clinical and genetic data graded the 80 test volumes.

Volume-level accuracy and weighted F1-score.

On the independent test set (80 volumes: 33 CHM, 29 USH2A, 18 controls), all models achieved high performance (accuracy ≥0.963). The MoE achieved the highest accuracy on the internal test set (accuracy 0.988; 79/80; weighted F1-score 0.988), followed by CNN + vote (accuracy 0.975; 78/80; weighted F1-score 0.975) and the hybrid CNN-Transformer (accuracy 0.963; 77/80; weighted F1-score 0.962). Ophthalmologists achieved lower accuracy (0.875 and 0.863). External validation on an independent Rennes University Hospital data set (n = 28 volumes; CHM = 6, controls = 10, USH2A = 12), processed identically without retraining, yielded accuracy 28/28 for CNN + vote, whereas the MoE and hybrid CNN-Transformer each achieved 26/28.

Deep learning applied to macular OCT volumes accurately distinguished CHM, USH2A-associated rod-cone dystrophy, and healthy controls, and achieved higher observed accuracy than masked human readers in this OCT-only setting. Both learned slice weighting and simple aggregation achieved strong performance, suggesting that increased architectural complexity may not be necessary for this 3-class classification task. Further validation in larger, more diverse, and clinically representative cohorts is required to determine the potential role of such models as complementary diagnostic support.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.