Medical AI · asia
AI Interprets Retinal Images, Helping Physicians Narrow the Genetic Scope of Inherited Retinal Diseases
A 300-person, multicenter randomized trial showed that Retina4IRD increased specialists’ top-five gene diagnosis accuracy from 67.3% to 88.5%. Its role is not to replace sequencing, but to provide more focused guidance before costly and complex genetic testing.
Inherited retinal diseases often begin with night blindness, narrowing of the visual field, or progressive vision loss, but similar retinal findings may be caused by different genes. Physicians must piece together imaging, medical history, and genetic information to approach the true cause. A randomized trial published in *Nature Medicine* showed that the artificial intelligence decision-support system Retina4IRD can help retinal specialists rank potential genetic diagnoses more accurately.
Retina4IRD is built around a vision transformer and integrates color fundus photographs, optical coherence tomography (OCT), and descriptive clinical data. It classifies cases into 17 genotype categories and outputs candidate results ranked by likelihood, along with image attention heatmaps. These categories also cover some mutations for which gene therapies are already available or are entering clinical trials, so more accurate preliminary triage may affect subsequent testing, referrals, and treatment evaluation.
The research team first trained and validated the system using data from 1,843 genetically confirmed patients from China, South Korea, and Poland, comprising 3,376 eyes. The model’s top-five prediction accuracy was 90.4% in internal validation and 85.6% in external validation; the latter better reflects the potential decline in performance when the model leaves its original data environment.
The more critical test came from a randomized controlled trial conducted at seven centers in China. A total of 300 patients with suspected inherited retinal diseases were assigned one-to-one to interpretation by physicians alone or interpretation with Retina4IRD assistance. Of these, 295 were included in the final analysis because they had analyzable next-generation sequencing reports, with the sequencing results serving as the reference standard for pathogenic mutations. Top-five gene diagnosis accuracy reached 88.5% in the AI-assisted group, higher than the control group’s 67.3%. When only the top-ranked answer was considered, the rates were 37.8% and 22.4%, respectively.
These results indicate that the system’s most practical value may not be identifying the cause correctly on the first attempt, but narrowing the search range. A post hoc analysis of the study also showed that the AI-assisted group had a higher composite score for subsequent management, but this type of analysis does not constitute independent efficacy evidence replacing the primary endpoint, nor can it prove that patients ultimately achieved better vision, quality of life, or treatment outcomes.
Retina4IRD should currently be understood as a clinical aid used before genetic testing, rather than a diagnostic device: candidate results must still be confirmed through formal genetic testing, variant interpretation, and genetic counseling. The trial was concentrated in Chinese medical centers. Although the training data included patients from three countries, prospective validation is still needed across more populations, different equipment, and general clinical settings. Before deployment in routine care, issues including data governance, liability for errors, model updates, and medical device regulation must also be addressed.