Models assist; evidence decides
Free-text sequence names vary by hospital and scanner. A combined system uses approved DICOM metadata, geometry and specialized classifiers to identify likely sequence or reconstruction types. Confidence and supporting fields remain visible so ambiguous cases can be reviewed.
Intelligence modules by programme
The model layer is modular rather than a single universal classifier. Each clinical programme activates the smallest set of models needed for its modality, anatomy and release criteria.
- MRI sequence and orientation classification
- MRE wave, elastogram and stiffness-map recognition
- CT reconstruction and anatomical-coverage checks
- Image corruption, duplication and geometry anomaly detection
- Report structuring and candidate-label extraction
- Dataset distribution and outlier analysis
Clinical truth remains distinct
A model can propose a series type or candidate label. It does not create definitive diagnosis or expert segmentation by itself. The platform keeps automated inference separate from clinician-validated ground truth and records the provenance of both.
Questions, answered directly.
Is one model used for every modality?+
No. Imaging tasks differ by modality, anatomy, protocol and output, so models are selected and versioned by programme.
Can model-generated labels be delivered as ground truth?+
Only if the agreed programme explicitly defines and validates that method. Automated candidates are otherwise kept separate from clinician-verified labels.
Why combine rules and models?+
Rules provide transparent deterministic checks; models help resolve variation and patterns that are difficult to encode reliably.