Medical imaging AI
Last reviewed: 2026-06-28 — see the freshness policy.
Learning objectives
After this chapter you will be able to:
- Describe the medical-imaging AI pipeline from DICOM to inference to clinical workflow.
- Place MONAI and the managed imaging services in an architecture.
- Account for the regulatory and integration realities of imaging AI.
The opportunity and the constraint
Medical imaging (radiology, pathology, ophthalmology) is the most mature clinical AI domain — most FDA-cleared AI devices are imaging. The data is DICOM; the models are deep neural networks; the constraint is that imaging AI used for diagnosis is usually a regulated medical device and must fit into the radiologist's existing workflow (PACS, worklist) to be adopted.
flowchart LR
PACS["PACS / DICOM store"] -->|"QIDO/WADO-RS"| Pre["Preprocess<br/>(normalize, window)"]
Pre --> Model["Imaging model<br/>(MONAI / custom)"]
Model --> Out["Findings, segmentation,<br/>triage score"]
Out --> Report["Back to worklist /<br/>report (FHIR / DICOM-SR)"]
Building blocks
- MONAI — the open-source PyTorch framework for medical-imaging deep learning (training, transforms, model zoo); MONAI Deploy packages models as clinical inference apps. Co-developed by NVIDIA. The default framework for custom imaging AI.
- Managed imaging stores — AWS HealthImaging, GCP Cloud Healthcare API DICOM, Azure DICOM service provide DICOMweb access to feed pipelines (see capability map).
- GPU compute — training and inference run on NVIDIA GPUs (cloud or on-prem DGX); large 3D volumes are memory-hungry.
- Viewers — OHIF and PACS integrations surface AI output where radiologists already work.
Architecture realities
- Workflow integration is the adoption gate. An accurate model that radiologists must leave their PACS to use will not be used. Results must return to the worklist (prioritize a likely-positive study) or the report (as a DICOM Structured Report or FHIR
Observation/DiagnosticReport). - Data volume & tiering. Studies are large; keep compute co-located with the imaging store and tier cold studies (see DICOM and cost).
- De-identification is two-layer. Imaging AI training data must scrub both DICOM tags and burned-in pixel annotations (see de-identification).
- Edge/real-time. For intra-procedure or device-attached inference, Holoscan runs at the edge with low latency.
Regulatory reality
Imaging AI that informs diagnosis is typically SaMD. Design for it from the start: define the intended use, lock or PCCP-manage the model, log inputs/outputs for traceability, and plan the validation evidence. A research prototype and a cleared product are very different builds — know which you are making.
Lab
A MONAI segmentation/classification model on a public dataset (e.g. TCIA) deployed against a local DICOM store is a strong extension lab (see the bootcamp lab model).
Check yourself
- Why is workflow integration (worklist/report) the make-or-break factor for imaging-AI adoption?
- What two layers must de-identification address for imaging training data?
- When does an imaging model become a regulated medical device, and what does that imply for the architecture?