Machine learning for medical imaging is software that learns from labeled scans — X-ray, CT, MRI, ultrasound, mammography — and returns a mark, a mask, or a score a clinician then uses or ignores. It is not a robot radiologist. It does not sign the report.
The same job gets sold as “machine learning in medical imaging,” “machine learning in radiology,” “machine learning medical imaging,” and a pile of numbered twins. Those are keyword wraps of this page. A 2023 news stamp is not a different article. A thin “what is MONAI” overview is not a tutorial. AI in radiology is the product job (second reader, worklist, signed report). This page is the method layer under that: how models learn from pixels. AI for medical diagnosis is CAD as a specialty. AI in healthcare is the broad stack. Computer vision in healthcare is the pixel-reading job across specialties, not just imaging.
What it actually does
Computer-aided detection has been around for decades: a hand-built detector that marks a candidate. Deep learning replaced most of those detectors with a convolutional net — sometimes a transformer — trained on labeled studies. The job did not change. After the exam, images go into a model. The model flags candidates. The clinician reviews the flags with the originals and signs.
Three outputs show up again and again:
- Classification. One label for the study or the region: “likely ICH,” “referable diabetic retinopathy,” “this mammogram is BI-RADS 4.”
- Detection. A box or a worklist bump: nodule on chest CT, bleed on non-contrast head CT, mass on a screening mammogram.
- Segmentation. A pixel mask: organ or lesion outline, tumor bed, vessel tree. The number you then compute (volume, diameter, doubling time) is a measurement, not a diagnosis.
| Area | Typical input | Typical output |
|---|---|---|
| Oncology | CT, MRI, PET, mammography | Lesion candidates, tumor volume, response on follow-up |
| Neurology | Brain MRI, non-contrast head CT | Bleed / infarct flags, lesion or atrophy maps |
| Cardiology | Cardiac CT/MRI, echo | Chamber / vessel measurements, function scores |
| Screening | Chest X-ray, mammography, low-dose CT | Second-reader marks, worklist bump |
Where the models actually run
Oncology. A net trained on labeled CTs or mammograms flags a lesion, outlines a tumor bed, or compares volumes across visits. The oncologist still treats. The radiologist still signs. A volume change is a measurement. It is not a complete response.
Neurology. Head CT for bleed or infarct. Brain MRI for lesion load or atrophy. The flags are candidates. Stroke and neurodegeneration work still belongs to the clinician who reads the study and the team that acts on it.
Cardiology. Chamber size, ejection fraction, vessel measurements on CT, MRI, or echo. The number is useful when it is re-validated on your machines. It is not a cath-lab decision by itself.
Screening. Second reader on mammography or low-dose chest CT, or a worklist bump on a chest X-ray. Most cleared imaging ML is assistive (FDA SaMD), not autonomous. A mark is a candidate. The reader still signs.
How an imaging ML system is built
The “machine learning medical imaging” posts that used to live on this site were this explainer with the preposition stripped. The pipeline is shorter than the listicles made it sound.
- Acquire. Pull DICOM from PACS, or NIfTI from a research folder. De-identify before anything leaves the premises.
- Preprocess. Intensity, spacing, orientation, crop. Medical volumes are not RGB photos. The MONAI how-tos are preprocessing 3D volumes and 3D augmentation — different jobs, real code.
- Train. A convolutional net or a transformer on labeled studies. Labels come from reports, biopsies, or expert reads — not from a vendor deck.
- Validate. A paper AUC is not a site number. Vendor, protocol, and population shift the score. Re-validate on your scanners.
- Deploy. Land the flags in the viewer the clinician already uses. A separate portal is how tools die. Training / evaluation / validation / deployment as their own articles sit in the ML lifecycle set, not here.
Radiology: the method layer
The old “machine learning in radiology” URL was this explainer with a different noun. Radiology as a product — second reader, worklist triage, who signs the report — is artificial intelligence in radiology. This page is the method under that: how a model learns from pixels and what it returns.
Will AI replace radiologists? No. The models are narrow. A lung-nodule net does not read a trauma pan-scan, does not take a call from the surgeon, and does not own the signed report. Volume grew; so did the need for people who can drop the false flags and keep the real ones. The longer version of that answer lives on the radiology page.
Python and MONAI
The old MONAI overview named the stack and stopped. MONAI is a PyTorch library for medical images (DICOM and NIfTI, not just RGB). Core is transforms and models. Labels is annotation. Deploy is packaging. That one-liner is enough for this page.
If you need the code, use the tutorials: preprocess, augment. If you need a clinic deployment, that is the MONAI / nnU-Net service. This page is not those tutorials. A 2023 “Python’s role” stub that named TensorFlow and stopped is also not a second URL.
What it does not do
- It does not replace the reader. The signed report is still a person. That is why most imaging ML is cleared as assistive, not autonomous.
- 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.
- 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 “$X billion by 2032” are not a reason to buy, and they are not this page.
- It is not a MONAI tutorial. Transforms, DataLoaders, and training loops live on the 3D preprocess and augment posts.
- It is not the radiology product page. Second reader, worklist, and “will AI replace radiologists” as a product story live on AI in radiology.
If the question is diagnosis / CAD as a specialty, that is AI for medical diagnosis. If the question is the broader healthcare stack, that is what is artificial intelligence in healthcare. If the question is computer vision across specialties (pathology, endoscopy, derm), that is computer vision in healthcare.
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.