Deep24‑2C predicts 10‑2 fields

Glaucoma clinics have a testing problem. The dense visual field test that best maps the central 10 degrees of vision — the Humphrey 10‑2 — is useful, but it takes another round in the bowl, more patient effort, and more fatigue. The standard 24‑2 grid is faster and more common, but it undersamples the center. (ophthalmologyglaucoma.org) Deep24‑2C is an attempt to bridge that gap by using AI to infer a full 10‑2 field from the newer 24‑2C test. (sciencedirect.com) ### What are these tests actually measuring? Visual field testing checks how well a patient sees dim lights at fixed spots. In glaucoma, those sensitivity losses form patterns, and the pattern matters as much as the average score. The 10‑2 test samples the central field densely. (aaojournal.org) The 24‑2 spreads points farther apart. The 24‑2C is a hybrid — basically the regular 24‑2 plus 10 extra central points added to catch damage near fixation. ### Why isn’t 24‑2C enough on its own? Because “more central points” is not the same thing as a full central map. A 10‑2 test measures 68 threshold locations in the central field. Deep24‑2C was built to predict all of them from just 22–26 24‑2C points that fall inside that same 10‑degree region, plus the patient’s age. (ophthalmologyglaucoma.org) That is the whole trick — can a sparse sample stand in for a dense one? ### What changed in this paper? The paper describes the development and validation of Deep24‑2C. The team trained several machine-learning models — Random Forest, XGBoost, and a multilayer perceptron — and also tested an ensemble that combined them. The goal was not just to flag “maybe abnormal,” but to reconstruct the point-by-point threshold values a real 10‑2 exam would have produced. (aaojournal.org) ### Why is central vision the hard part? Because central defects matter a lot to daily life — reading, faces, contrast, fine work — and glaucoma can damage that area earlier than older testing habits assumed. Earlier work showed standard 24‑2 fields can miss central defects that show up on 10‑2 testing. (sciencedirect.com) That is why clinicians often end up ordering both. ### So is this replacing the 10‑2? Not really — at least not yet. Think of it more like a synthetic preview. If the predicted 10‑2 looks clean, a clinic might skip an immediate second test. If the prediction shows suspicious central loss, that patient can be sent for the real 10‑2 or followed more closely. The value is triage and workflow, not a magic exemption from confirmatory testing. (ophthalmologyglaucoma.org) That triage use is an inference from the model’s design and the clinical gap it targets. ### What’s the catch? Visual fields are noisy. Patients get tired, fixation drifts, and thresholds bounce around from test to test. Any model trained on one test to predict another inherits that mess. (aaojournal.org) And even if the average prediction looks good, glaucoma decisions often hinge on small local defects near fixation, where a miss matters. So the bar is not “pretty similar.” The bar is “similar enough that you would trust the next decision.” ### Why are people paying attention? Because this is the kind of AI use case medicine actually needs — not replacing the doctor, just removing redundant work. If a single faster test can reliably approximate a second one, clinics save time, patients do less repetitive testing, and central damage might still get surfaced earlier than it would with a plain 24‑2 alone. (ophthalmologyglaucoma.org) ### Bottom line? Deep24‑2C is a practical idea: use the extra central information already present in 24‑2C to estimate the denser 10‑2 map. If the validation holds up in broader real-world use, the payoff is simple — fewer extra field tests for some patients, and faster identification of the ones who truly need them. (aaojournal.org) (ophthalmologyglaucoma.org)