Can AI read medical scans?
Sometimes, on a narrow measured task: pattern matching on pixels, not a radiologist in software: with the DeGrave shortcut paper as the honest door into what can go wrong.
Primary source: DeGrave, Janizek & Lee, Nature Machine Intelligence (2021): COVID CXR models that selected shortcuts over disease signal: https://doi.org/10.1038/s42256-021-00338-7
Headlines love a simple story: the AI “read” the X-ray. Patients hear a digital radiologist. Clinicians hear a colleague who never sleeps.
The useful answer is narrower, and more honest.
Sometimes, on a narrow measured task, a model can flag or score something a dataset defined. That is not the same as reading a scan the way a radiologist does: history, symptoms, priors, doubt, and responsibility included.
What the model actually does
Under the hood this is pattern matching on pixels (or voxels, the 3D cousins of pixels in CT and MRI). Someone gathered labeled examples. Someone trained a model so its guesses got closer to those labels. Later it outputs a score or a box on a new image.
That is the same practice loop from How a model learns: guess, measure error, nudge the weights. The curriculum is the pile of scans and labels, see What training data means. Chatbots are a different product shape on different data. Medical scan models are not “ChatGPT looking at your chest.”
Nothing here is conscious. Nothing here “understands” you as a person. It predicts from patterns that survived training.
Where it can help
Used carefully, scan models can be useful tools:
- Second reader, a second pass that catches something a tired eye might miss, with a human still in charge.
- Triage, sorting a worklist so urgent-looking cases rise sooner (still reviewed by people).
- Flagging a specific finding the training set defined, a nodule candidate, a fracture category, a pneumonia-like pattern, not “everything wrong with this patient.”
Help means assist. It does not mean replace the clinician, explain why you feel short of breath, or decide your care plan. This site does not give medical advice; neither should a scan score.
What it cannot do
- Replace judgment. Labels are not a full consult. Edge cases, rare diseases, and odd presentations are exactly where “good average accuracy” can fail.
- Generalize by default. A model trained at one hospital on one scanner vendor can stumble at another site, different machines, protocols, patient mix, even how markers are placed on the film.
- Understand the patient. It does not know your story. It does not owe you an explanation in clinical language unless someone built that on purpose, and even then, check it.
If a demo looks perfect on one public dataset, ask what world that dataset is.
Shortcuts: when high scores hide the wrong cue
Here is the honest technical door for this piece.
In 2021, Alex DeGrave, Joseph Janizek, and Su-In Lee showed that some deep learning systems for COVID-19 detection on chest radiographs looked strong while leaning on shortcuts, easy image cues that correlated with labels but were not the disease. Think laterality tokens (L/R markers), style differences between data sources, and other confounds that travel with “positive” vs “negative” piles. When the setting changed, the “signal” the model had learned could stop being medical.
That is the same family of failure as the watermark example already on this site: the model seems to detect disease but is really reacting to a hospital mark, a token in the corner, or a dataset-construction quirk. High numbers on the original test can be a cheat code, not clinical competence.
AUROC (a common ranking score for “can this separate positive from negative?”) on one dataset is not a clinic. You also need held-out sites, honest operating points (where you set the yes/no threshold), and what false positives cost per scan in a real workflow.
What papers here actually claim
Not every medical AI brief on Bitware is “the model stared at pixels.”
- GRIPNet nodule detector, a methods note about detecting lung nodules in CT with a design tied to how nodules look. Read it as a technical recipe with benchmarks, not a care recommendation.
- SIFTING lung-cancer T-staging, extracts staging structure from radiology reports (text), with links back to source phrases. Useful contrast: not all medical AI is scan-reading. Some of it is careful document work.
Browse those for what the papers claim, and for the caveats each brief already puts up front.
For the ImageNet-era roots of modern vision models that made scan AI practical at scale, the History door is The Deep Learning Turn.
Easy → technical doors
| If you’re curious about… | Easy / History next step | Technical door |
|---|---|---|
| Guess → error → update | How a model learns | Sources linked in that post |
| Labels, piles, watermark shortcuts | What training data means | Datasheets for Datasets via that post |
| Pattern engines vs. movie AI | What AI actually is | Sources linked there |
| Shortcuts over disease signal | This piece’s primary source | DeGrave et al., Nat Mach Intell 2021 |
| Nodule detection (pixels) | GRIPNet | Paper linked in that brief |
| Staging from reports (text) | SIFTING | Paper linked in that brief |
| Vision’s deep-learning turn | The Deep Learning Turn | ImageNet / AlexNet trail via History |
The honest caveat
Scan AI can be a measured assistant on a defined finding. It is not a radiologist, not a diagnosis, and not permission to skip human review. Good systems name the task, the data, the failure modes, and the operating point. Bad marketing says “the AI read your scan” and stops talking. On this site, Easy reading aims for the first kind of story, clear claims, clear limits, and a path to the paper when you want to verify.