AI in medical images is software that marks, segments, or scores a scan so a clinician can use or ignore the mark. This page is why those jobs exist — de-identify, segment a vertebra, color a tumor class — not a definition of healthcare AI and not a diagnosis explainer.
If you meant what AI in healthcare is → what is artificial intelligence in healthcare. If you meant diagnosis / CAD → AI for medical diagnosis. If you meant the de-id how-to → DICOM anonymization using Python / DICOM anonymizer software. If you meant the label that trains the model → medical image annotation.
PYCAD builds custom web DICOM viewers and imaging models (annotation → deployment). It does not ship a diagnosis product, a public annotation app, or a de-id SaaS. The old “free spine annotator” and “anonymization web app” links on this URL 404; they are not products.
Why these jobs exist
A blog that only shows how to load a NIfTI is answering the engineer. A clinic is asking a different question: what does that pipeline change on the worklist. Four jobs show up again and again. None of them is “AI diagnoses the patient.”
| Job | What the model returns | Why anyone bothers | Who signs |
|---|---|---|---|
| De-identify | A copy of the study with names, IDs, and burned-in text gone | You cannot train or share the file while it still names the patient. HIPAA is the constraint, not a badge. | The privacy officer / the protocol. Not a model. |
| Segment an organ | A mask: this voxel is vertebra / lung / liver | A measurement, a plan, a print. Without the mask you are eyeballing slices. | The radiologist or the surgeon using the mask. |
| Detect / classify a lesion | A box, a score, or a color per class (tumor core / edema / necrosis) | The worklist flag, the volume over time, the target for radiation. | The radiologist. The color is a suggestion. |
| Screen a population task | A refer / don’t-refer on a standard photo (retina, skin, a screening MRI) | The specialist is scarce. The camera is not. | Usually a clinician. A few cleared systems are autonomous for one narrow label — that is the device page, not this one. |
De-identify first
A CT that still has the patient’s name in a tag, or burned into a pixel, is not a training set. It is a disclosure. Anonymization strips the identifiers so the pixels can leave the premises — research, a vendor, a model. It does not make the image “safe forever.” Dates, unusual studies, and faces in a reconstruction can still re-identify. The how-to is the Python job and what data anonymization is. This page only names why the step exists: no de-id, no legal training data.
Why segment a vertebra
A fracture that is easy to miss on a scroll is easier to argue about once the vertebral body is a mask. The mask is also how you measure height loss, plan a screw, or print a model. Same job on lung or liver: the organ outline is the input to everything after. A Slicer + nnU-Net how-to for spine already lives at automatic spine segmentation in 3D Slicer. This is not that tutorial. This is why the mask is worth making.
Why color a brain tumor
Glioma work is not “tumor / no tumor.” BraTS-style labels split enhancing core, edema, and necrosis because the surgeon, the radiation plan, and the follow-up volume each need a different outline. Multi-class color is a map, not a diagnosis. The report is still a person. Detection-as-diagnosis is AI for medical diagnosis.
Three screening examples (same pattern)
- Retina. A fundus photo, a refer-if-more-than-mild-diabetic-retinopathy flag. One of the few tasks with an autonomous cleared device (LumineticsCore, formerly IDx-DR). Named devices sit on artificial intelligence medical devices.
- Skin. A dermoscopic photo, a priority for biopsy. The model ranks; the dermatologist still cuts or doesn’t.
- Brain screening. PET or MRI scored for patterns associated with Alzheimer’s-type change. A risk flag is not a diagnosis of dementia. The neurologist still sees the patient.
Those three are the same second-reader pattern on a different pixel type. They are not a second article each.
What this page is not
- Not “what is AI in healthcare.” Stack, NLP, claims models, implementation path → 6151.
- Not CAD / diagnosis as a specialty topic → 5927.
- Not a patient story. The old “consider a patient with a subtle fracture” paragraphs were costume. Dropped.
- Not a YouTube or Medium-tag footer. Dropped.
If the work is a viewer or a model on those pixels, that is the imaging piece. Case studies.