Medical Technology · global
From Identifying Macular Holes to Writing Reports: Ophthalmic AI Achieves 98.8% Accuracy in External Testing
A proof-of-concept study linked retinal image interpretation and patient reports into a workflow, but its high accuracy came from a binary test of macular hole versus normal images, leaving a gap between the results and the complex cases seen in real clinical practice.
Findings from ophthalmic images must be translated into explanations patients can understand to complete an important step in clinical communication. A study published on October 23 in *Ophthalmology and Therapy* attempted to connect macular hole identification and report writing into an automated workflow, exploring how image analysis could feed into subsequent communication.
The study used optical coherence tomography (OCT) images as input. A fine-tuned retinal foundation model, RETFound, first classified images as showing either a macular hole or a normal macula, after which GPT-4 generated explanations. The language model did not view the images directly or participate in diagnostic decisions; image classification and text generation served distinct roles.
Across 1,014 external validation images, classification accuracy reached 98.8%, with sensitivity of 97.9%. Of 563 macular hole images, 12 were missed, while all 451 normal images were classified correctly. These figures refer to images and cannot be directly treated as clinical validation results for the same number of people.
The external validation also had a key limitation: the macular hole and normal images came from different datasets. It therefore remains impossible to fully rule out whether differences in scanning equipment, image acquisition methods, or populations partly influenced the model’s performance. The researchers noted that multicenter data containing both types of cases are needed to confirm generalizability.
The performance of the written reports was less clear-cut than the classification figures. Two ophthalmologists assessed 40 reports, with both judging 37 to be accurate and sufficient. Another reviewer noted that two reports underestimated how common cataract formation is as a surgical complication, while one omitted relevant surgical risks, showing that fluent text may still contain errors in medical information.
The model currently distinguishes only between macular hole and normal images. It has not yet been shown to distinguish other conditions with a similar appearance, nor has staging or prognosis been evaluated. It analyzes a single two-dimensional cross-section, which may miss small or off-center holes; the generated reports also did not incorporate individual clinical details such as visual acuity or symptom duration.
The significance of this proof of concept lies in demonstrating how image interpretation and explanations can be connected. Before use in actual care, prospective clinical validation is still needed to establish the reliability of the full workflow in complex cases and whether patients truly understand the reports. At this stage, final review by a physician remains necessary.