AI in Medicine Watch — Sep 17, 2026

Welcome to your weekly update in AI in Medicine.

This period is dominated by a single question, asked ten different ways: not whether a model can match an expert, but what happens to the human being sitting in front of the images when the model is switched on. We have a genuine randomized trial of decision support in ophthalmology, a reasoning language model reading radiology reports, a generalist vision-language model for abdominal CT, two fracture-detection papers that pull in noticeably different directions on generalization, three studies pushing past detection into risk and prognosis, and a pair of papers about where clinical information comes from in the first place. If you are deciding whether to deploy an assistive tool this year, this is a useful period.

Start with the expertise gap, because three papers attack it from very different evidentiary altitudes. The strongest is in Nature Medicine, where Jia and colleagues built Retina4IRD, a system that predicts seventeen genotype categories for inherited retinal disease from fundus photographs and optical coherence tomography. They trained it on nearly two thousand genetically confirmed patients across China, South Korea and Poland, and then — unusually — they ran an actual randomized controlled trial, assigning three hundred patients with suspected inherited retinal disease to specialist assessment with or without the tool, with next-generation sequencing as the reference standard. With assistance, the correct gene appeared in the specialist's top five predictions in nearly nine out of ten patients, against roughly two thirds without it. Management decisions improved too. The honest caveat is that the top single guess was still correct in fewer than four in ten cases even with the model, so this is a tool for narrowing a differential before genetic testing, not for replacing it — and the trial was conducted by specialists, in centres that already see this disease.

Compare that with Radiology, where Li and colleagues tested GPT-5-Thinking on narrative MRI reports of orbital and head-and-neck tumours across a generalist centre and a subspecialist centre, a thousand patients with pathologically confirmed diagnoses. The model, working only from the text of the report, matched subspecialists at roughly three quarters accuracy for orbital tumours. The more arresting finding is the gap it closed: routine generalist reports for head-and-neck tumours got the top diagnosis right in about a quarter of cases, and the model reached about two thirds. When seven generalists reread cases with the model's help, their accuracy rose roughly fourteen points on head-and-neck tumours. This is retrospective, convenience-sampled, and it evaluates language reasoning over descriptions someone else already wrote — so garbage descriptions will still yield garbage diagnoses. But it suggests something practical: a reasoning model applied at the report level may be a cheap intervention where subspecialty coverage is thin.

Then there is the scale play, in Science, where Zhang and colleagues present RADAR, a generalist vision-language model trained on more than four hundred thousand contrast-enhanced abdominal CT examinations, learning directly from clinical reports without manual annotation, and covering eighteen anatomical structures and one hundred forty-six findings. In a reader study, twenty-six radiologists gained about ten percent in sensitivity with the model's assistance. The trained-from-reports approach is the important part — it dissolves the annotation bottleneck that has capped supervised models. What we are not given is prospective deployment data or a clear account of specificity cost, and a model spanning one hundred forty-six findings has one hundred forty-six opportunities for a subtle failure mode. Taken together, these three agree that assistance lifts the less-specialized reader more than the specialist — and only the retinal trial proves it in a prospective, randomized setting. That asymmetry between trial-grade and reader-study evidence is the story of the period.

Fractures give us the period's clearest disagreement, and it is about generalization. Lin and colleagues, in Radiology, developed OccuNet for femoral-neck fractures on pelvic or hip radiographs, using contrastive pretraining on artifact-augmented images, across four hospitals and more than two and a half thousand patients with same-episode CT or MRI as reference. Performance was near-ceiling — sensitivity and specificity both around ninety-eight percent — and, critically, for radiograph-negative or indeterminate fractures the model found about ninety-five percent, against eighty-six percent for musculoskeletal radiologists and under seventy percent for emergency physicians. With assistance, emergency physician sensitivity rose to about ninety-six percent and reading times fell by roughly a sixth.

Now hold that against the multinational validation from Ruitenbeek and colleagues, also in Radiology, which took a pretrained appendicular fracture tool into three European hospitals, fifteen hundred patients, with local expert radiologists setting the reference standard. Here sensitivity ran between eighty and eighty-six percent by site — respectable, but nowhere near the femoral-neck figures — and wrist, hand and finger examinations at one hospital dropped to about seventy-five percent while the other sites sat above ninety. Reader sensitivity still improved by about eleven percentage points with no specificity penalty, which is the reassuring part.

Say the disagreement plainly: one study says fracture AI is essentially solved, the other says performance is anatomy-dependent and site-dependent even when overall figures look stable. Both can be true — a narrow, single-anatomy model with a CT or MRI reference standard is a fundamentally easier and better-adjudicated task than a broad appendicular tool judged against local radiologist consensus. What would settle it is OccuNet evaluated prospectively outside its development health system, with the same advanced-imaging reference, and reported by anatomy and by site rather than pooled. Until then, treat single-site ceiling numbers as a hypothesis, and demand your vendor's per-site, per-anatomy breakdown before go-live.

Three papers move past yes-or-no detection into grading, prognosis and trust. Song and colleagues, in Radiology, built a radiograph-based deep learning model for bone tumour risk stratification across ten institutions, with a prospective test cohort of one hundred fifty-two participants. It substantially outperformed both nine musculoskeletal radiologists and the Bone Reporting and Data System, and with assistance readers gained discrimination and, notably, agreed with each other more. That improvement in interreader agreement may matter more in practice than the accuracy gain, because it is variability that drives unnecessary biopsy. Alongside it, another Radiology study applied contrast-enhanced CT radiomics to more than eleven hundred patients with clinical stage one lung adenocarcinoma to predict high-grade histologic patterns preoperatively; the model-defined high-risk group had roughly three times the recurrence risk, validated externally. That is prognostically real, but it is retrospective, surgical-cohort-derived, and radiomic features remain notoriously fragile across scanners — do not let it change the extent of resection yet.

The trust question is tackled head-on in PLOS Medicine, where Thrun and colleagues developed FlowXAI for B-cell non-Hodgkin lymphoma classification from flow cytometry, evaluated on nearly twenty thousand peripheral blood samples plus an external dataset using a different antibody panel. It matched a deep learning system while needing roughly a hundredfold less training data, and when it flagged its own prediction as confident, it beat that neural network baseline. It also said honestly which entities its panels could not separate. Self-assessed confidence is the feature to want in any diagnostic tool you adopt — but this remains retrospective and panel-specific, and needs prospective workflow validation.

Finally, two papers about what information reaches the clinician at all. In Nature, Liao and colleagues describe passive heart-rate monitoring using facial video photoplethysmography during ordinary smartphone use — developed and validated on over three hundred fifty thousand videos, with error under ten percent against electrocardiogram across light, medium and dark skin tones, and daily resting heart rate within five beats per minute of a wearable. That is a cardiovascular biomarker without a device, and the explicit skin-tone equity testing sets a standard others should meet; the population was small in participant numbers, and no outcome benefit has been shown. And in Annals of Internal Medicine, Lea and Podolsky trace the history of navigating medical literature, from Billings' indexing through abstract journals and citation indices to computerized retrieval, arguing that search systems are not neutral maps of knowledge but constitute it — shaping journal rankings, research priorities and practice. Read it as a warning label for the language model now summarizing your evidence.

If you read only one paper from this period, make it the Retina4IRD trial in Nature Medicine. It is the rare randomized, prospective demonstration that decision support changes not just diagnostic accuracy but downstream management — and it sets the evidentiary bar every reader study in this briefing should be measured against.

Here is what this period adds up to. First, assistance reliably lifts the less-specialized reader, and that is now the most robust finding in clinical AI. Second, the evidence base is still overwhelmingly retrospective reader studies; one randomized trial does not make a field. Third, generalization is unsettled, and the fracture papers show it — pooled performance can hide anatomy-specific holes. Fourth, watch for models that report their own uncertainty and that publish subgroup performance by site and by skin tone, because those are becoming the markers of deployable tools rather than publishable ones. And finally, keep one eye on the knowledge layer: the same technology reading your images is increasingly deciding which literature you ever see.

That's your AI in Medicine update for this period. Until next time.

This is an automated summary generated by artificial intelligence, which can make mistakes. Always review the original source materials.