Why this study matters

Choroideremia (CHM) is an inherited retinal disease, and accurately identifying it is an essential part of understanding a person’s condition. Yet retinal disorders can share features on imaging, particularly when they cause rod-cone degeneration. This can make image-based interpretation challenging, even for experienced clinicians.

A 2026 study in Ophthalmology Science explored whether artificial intelligence (AI) could help distinguish choroideremia from USH2A-associated rod-cone dystrophy and from healthy retinas using macular optical coherence tomography (OCT) volumes. OCT is a noninvasive imaging technique that creates detailed cross-sectional views of retinal layers. For patients and families, this research is important because it examines whether routinely acquired imaging could eventually provide useful additional diagnostic support alongside clinical assessment and genetic testing.

How the researchers studied OCT volumes

The team conducted a retrospective diagnostic study using 535 macular OCT volumes from 262 participants. The dataset included 99 people with choroideremia, 96 with USH2A-associated disease, and 67 healthy controls. All scans were acquired with a Heidelberg Spectralis device between 2012 and 2021.

An OCT volume is not a single picture: it is made up of multiple individual cross-sectional images, known as B-scans. To make the data consistent, the researchers standardized each volume to 19 B-scans. They then separated the scans at the patient level into training, validation, and test groups. This is important because scans from the same individual were not split across these groups, helping ensure that the final test measured performance on people the models had not already encountered.

The investigators compared three AI approaches:

  • A slice-based convolutional neural network (CNN) that assessed individual B-scans and used a majority vote to reach a volume-level result.
  • A Mixture-of-Experts (MoE) model, designed to learn which B-scans within a volume should carry the most weight.
  • A hybrid CNN-Transformer model, designed to combine image features from individual slices while considering information across the full OCT volume.

Two ophthalmologists also reviewed the test scans. They were masked to clinical and genetic information, meaning their decisions were based on OCT imaging alone.

Key findings: very high accuracy from OCT-based AI

On the independent internal test set of 80 OCT volumes—33 from people with choroideremia, 29 with USH2A-associated disease, and 18 from healthy controls—all three AI models performed strongly. Each achieved an accuracy of at least 96.3%.

The MoE model had the highest internal-test performance. It correctly classified 79 of 80 volumes, for an accuracy of 98.8%. The CNN with majority-vote aggregation correctly classified 78 of 80 volumes, an accuracy of 97.5%. The hybrid CNN-Transformer correctly classified 77 of 80 volumes, for 96.3% accuracy.

The researchers also reported weighted F1-scores, a measure that considers both correct identification of each group and the balance of errors across groups. These scores closely matched the accuracy results: 0.988 for the MoE model, 0.975 for CNN plus voting, and 0.962 for the hybrid model.

In this OCT-only comparison, the two masked ophthalmologists achieved accuracies of 87.5% and 86.3%. This does not mean AI replaces specialist care. Rather, it shows that the models recognized imaging patterns that distinguished these three groups very effectively under the study conditions.

External validation strengthens the result

A major strength of the work was external validation. The models were tested without retraining on 28 OCT volumes from an independent dataset at Rennes University Hospital: six CHM volumes, 12 USH2A-associated disease volumes, and 10 control volumes.

The simpler CNN-plus-vote approach correctly classified all 28 external volumes. The MoE and hybrid CNN-Transformer models each correctly classified 26 of 28. This finding is notable because the models retained high performance on scans from a separate dataset rather than only on data from the original study population.

It also highlights an important practical lesson: more complex AI architecture was not necessarily better for this three-class task. Both learned weighting of individual scan slices and straightforward majority-vote aggregation produced strong results.

What this could mean for diagnosis and treatment pathways

The study points toward a possible future role for AI as complementary diagnostic support. In a clinical setting, an OCT-based model could potentially help flag imaging patterns that are more consistent with CHM or USH2A-associated rod-cone dystrophy. Such support could be especially valuable when specialists are interpreting large numbers of imaging slices.

For families navigating inherited retinal disease, correctly distinguishing conditions matters because CHM and USH2A-associated disease are genetically distinct. However, this study evaluated OCT images only. The authors emphasize that further validation is needed in larger, more diverse, and more clinically representative groups before the potential clinical role of these tools can be determined.

Looking ahead

This research demonstrates that macular OCT volumes contain highly informative signals for separating choroideremia, USH2A-associated rod-cone dystrophy, and healthy retinal anatomy. It also shows the value of testing AI systems on independent data from another hospital.

The next steps will be to evaluate these models across broader populations and real-world clinical settings. If their performance remains strong, AI-assisted OCT interpretation may become a useful addition to the diagnostic toolkit—supporting, rather than replacing, the combined expertise of retinal specialists, clinical assessment, and genetic evaluation.