MPIML’s missing bridge from OCT to the phone

The measurement the phone must reach

Estimating retinal-layer thickness from a fundus photograph is a worthwhile research goal where access to expensive imaging equipment is limited. MPIML, a machine-learning approach published on 26 October in the research journal Scientific Reports, pursues that goal. Its name describes a framework combining multiple imaging modalities with physical constraints. The authors aim to learn from optical coherence tomography (OCT), which produces cross-sectional retinal images, and use only a phone-acquired fundus photograph for a new patient. The photograph supplies a surface view; OCT supplies depth. The interesting question is how the model connects those surface cues to thickness in the cross-section. That connection is my focus in a study the authors describe as a 23-patient feasibility pilot.^[1]

The camera is the first link needing clarification. The title and abstract foreground smartphone images, while the dataset section describes fundus photographs acquired with a TOPCON camera integrated into an OCT platform. The main article does not clearly document a separate phone cohort testing the same prediction or explain how transfer between devices was evaluated. That matters at the centre of the accessibility proposal: taking a learned relationship to someone using a phone requires knowing the actual acquisition conditions. A method developed with one camera family may work with another. What remains missing here is the experimental bridge identifying the phone images and the comparison used to test that transfer.^[1]

OCT information inside the thickness estimate

The next link is the anatomical interval being predicted. The abstract and conclusion name the retinal nerve fibre layer, or RNFL. Yet the OCT segmentation method names the inner limiting membrane and retinal pigment epithelium, abbreviated ILM and RPE, as its boundaries. The comparison table also describes an ILM–RPE thickness measurement. Those different anatomical boundary names do not define an interchangeable target. Before interpreting a reported prediction error, the paper needs to explain which boundary pair generated the OCT reference thickness. My objection concerns the unestablished match between that training target and the measurement expected from a phone; it does not establish that prediction from a photograph is impossible.^[1]

The network inputs raise another question. The method defines ganglion cell complex (GCC) thickness as an OCT-derived measurement of inner retinal layers. Its structural physics index uses the ratio of GCC to RNFL; other indices relate vessel density to GCC and optic-disc cupping to RNFL. These indices also appear in the final network input representation. If a new patient supplies only a photograph, the route to these OCT-dependent quantities needs explanation. They might be confined to training, or generated through intermediate estimates from the photograph. The main article does not explain that distinction adequately in its input chain. This uncertainty does not prove that test information leaked into training; it calls for an explicit account of computation without OCT.^[1]

The link validation has not completed

The safeguards described by the authors deserve weight in this assessment. They state that patients were separated before image augmentation, that augmentation occurred only within training folds, and that feature scaling used values calculated from training data. These are important design precautions. But the working cohort selects one eye per patient and lists 16 training patients, 3 validation patients and 4 held-out test patients, while the thickness-results narrative refers to 8 test eyes. The main article does not reconcile which evaluations these counts describe. That prevents a straightforward reading of the results as one patient-level test. Both a clear account of safeguards and a clear identification of the eyes used for prediction evaluation are necessary.^[1]

The external-validation table makes the remaining distinction concrete: it covers fundus components, while OCT-referenced RNFL thickness regression remains unvalidated. Locating structures correctly in a photograph and predicting cross-sectional thickness from that photograph have different success criteria. The next step that would persuade me is an evaluation that clarifies the camera and anatomical target and specifies how OCT-dependent inputs are obtained for a new patient. Comparing paired phone photographs and OCT measurements from patients untouched by model development directly tests the proposed transfer. The access problem remains a valuable reason for the research. Progress beyond this pilot requires making that particular passage from image to thickness open to examination.^[1]