Imaging AI
Medical imaging was the first area of healthcare where machine learning reached routine clinical deployment. Images are standardized, plentiful and already digital — and radiology workflow has a well-defined place to insert a result.
DICOM foundations
DICOM (Digital Imaging and Communications in Medicine) is both a file format and a network protocol. Anyone working with imaging data works with it.
Hierarchy: Patient → Study → Series → Instance. A study is one imaging examination; a series is one acquisition within it; an instance is usually one image.
Key services:
- C-STORE / C-FIND / C-MOVE — classic DICOM networking
- DICOMweb — the REST equivalent:
WADO-RS,QIDO-RS,STOW-RS - PACS — the archive where studies live
- Modality worklist — how scanners know what to acquire
Pixel data carries a large amount of embedded metadata, including patient identifiers — which is why de-identification is not just a matter of blanking a filename.
Task types
| Task | Example |
|---|---|
| Classification | Is this chest X-ray abnormal? |
| Detection | Where is the nodule? |
| Segmentation | Outline the organ or lesion |
| Quantification | Volume, density, growth over time |
| Triage | Which studies should be read first? |
| Quality control | Was this acquisition adequate? |
Screening use cases in settings with few radiologists — tuberculosis screening on chest X-ray, diabetic retinopathy screening on fundus images — are where the access argument is strongest.
Evaluation pitfalls
Shortcut learning. Models learn scanner artefacts, laterality markers, drain tubes or text burned into the image rather than pathology. External validation on different equipment is the only reliable check.
Site leakage. Splitting by image instead of by patient and site inflates scores.
Label noise. Radiology reports are the usual label source and they are free text with hedging. NLP-extracted labels carry the extraction error into the model.
Selection bias. Studies that were ordered are not a random sample of the population.
Prevalence dependence. Positive predictive value depends on prevalence; performance reported on an enriched dataset does not transfer to screening.
Workflow integration
An imaging model is useful only if its output arrives where the radiologist works:
- Results returned as DICOM Structured Report or secondary capture, or
as FHIR
Observation/ImagingStudyresources - Worklist prioritisation for triage use cases
- Turnaround time compatible with reading workflow
- A clear indication that the finding is machine-generated
- Audit of what was suggested and what the reader decided
Clinical AI regulation applies in full — see clinical AI.
Related
- Machine learning
- AI ethics
- GCP health, AWS health, Azure health — managed DICOM stores