AI for medical diagnosis is software that reads images, labs, notes, or waveforms and flags findings a clinician then accepts or rejects. It does not write the report. It does not replace the diagnosis.
In radiology the same job has been called computer-aided diagnosis (CAD) for decades: after the exam, images go into the system, the system marks candidates, and the radiologist reviews the marks with the originals. CAD is a name for this workflow, not a second topic. Detection (finding candidates) is not diagnosis (the signed report). Mammography CAD is a different article — this page is the generic explainer.
How it works
- Train. A model sees labeled studies — a chest X-ray plus “nodule / no nodule,” a whole-slide image plus a tumor grade, an ECG plus a rhythm. The labels come from reports, biopsies, or expert reads. Garbage labels produce garbage flags.
- Infer. On a new study the model outputs marks, scores, or a short list of differentials. That is a suggestion, not a chart note.
- Review. A radiologist, pathologist, or treating clinician keeps or drops each flag and signs the report. Most cleared imaging AI is assistive (FDA SaMD), not autonomous.
The stack is ordinary machine learning. Convolutional nets dominate pixels (X-ray, CT, MRI, OCT, slides). NLP reads notes and labs. Tabular models sit on vitals and labs for triage. “Deep learning” is the image-era name for the same job; “machine learning” is the broader one. None of those are a different product.
Where the name came from
Rule-based consult systems tried this first. INTERNIST-1 (University of Pittsburgh, 1971) matched internal-medicine findings to a knowledge base. MYCIN (Stanford, mid-1970s) recommended antibiotics for bacteremia from explicit rules. Later systems such as QMR and DXplain used Bayesian scores. None of them read pixels. Overview: MYCIN, INTERNIST-1.
The imaging wave is different: labeled pixels plus GPUs. That is why “AI diagnosis” in 2020s marketing almost always means a CAD-style second reader on a scan, not a chat bot writing a differential from symptoms.
| Era | Systems | Method | What they actually did |
|---|---|---|---|
| 1970s | INTERNIST-1, MYCIN | Rules / expert knowledge | Internal medicine consult; antibiotic choice |
| 1980s–1990s | QMR, DXplain | Bayesian scores | Differential from findings, not images |
| 2000s | Early CAD on PACS | Classical image features | Marks on mammography and chest CT |
| 2010s–now | Deep learning on pixels | CNNs (and later transformers) | Second reader on X-ray, CT, MRI, OCT, slides |
Where it is used
| Specialty | Input | Typical flag | Who signs |
|---|---|---|---|
| Radiology | X-ray, CT, MRI, mammography | Nodule, bleed, fracture, mass | Radiologist |
| Pathology | Whole-slide images | Tumor, mets, grade / biomarker | Pathologist |
| Cardiology | ECG, echo | Arrhythmia, wall-motion | Cardiologist |
| Ophthalmology | OCT, fundus | Referral-warranted retinal disease | Ophthalmologist |
| Emergency / triage | Worklist + vitals | Priority, deterioration risk | Treating team |
Two published image results that are actually cited, not market decks:
- De Fauw et al., Nature Medicine 2018 — a deep-learning OCT referral system matched expert ophthalmologists across dozens of retinal conditions. Paper: doi:10.1038/s41591-018-0237-x.
- McKinney et al., Nature 2020 (Google Health) — a mammography model reduced false negatives 9.4% and false positives 5.7% versus typical readers in that study. Paper: doi:10.1038/s41586-019-1799-6. That is one screening task, not “AI diagnoses cancer.”
What it does not do
- It does not replace the reader. A mark is a candidate. The signed report is still a person. That is the CAD detection-vs-diagnosis split, and it is why most imaging AI is cleared as assistive.
- It fails on shift. A model trained on one vendor, protocol, or hospital will miss or over-call on another. You re-validate on your scanners; you do not copy a paper’s AUC onto the worklist.
- It inherits the labels. Narrow demographics, one site, or noisy reports become biased flags. That is a dataset problem, not a “fairness module.”
- It is not a market forecast. Vendor CAGRs and “92% of mammograms” slides are not a reason to buy, and they are not this page.
If the question is breast-screening CAD specifically, that is a different job: mammography with CAD.
PYCAD builds the imaging side of this — custom pipelines and web DICOM viewers when the study has to live in a clinic app. Case studies.