MedTech AI Diagnostic Imaging Platform — UK Radiology Group

MHRA SaMD Classification and Clinical Evidence: MHRA Class IIa: AI diagnostic triage that informs but does not replace radiologist decision. Clinical evaluation: systematic literature review of AI diagnostic performance evidence (AUC ≥ 0.92 required for pneumothorax detection). Intended purpose limitation: triage support only — AI highlights probable urgent findings to radiologist, does not output a diagnosis. Technical file: classification rationale, clinical evaluation report, post-market surveillance plan. AI Model Architecture: Chest X-ray classification model: EfficientNet-B4 pretrained on ImageNet, fine-tuned on CheXpert dataset (224,316 chest radiographs with radiologist labels). UK training data: MIMIC-CXR (MIT + Beth Israel) supplemented with 12,000 de-identified NHS chest X-rays under HRA research ethics approval. Output: 5 urgent findings probability scores (0–1) + triage tier (urgent/routine/normal). Inference: AWS SageMaker real-time endpoint (p95 latency < 200ms for 512×512 DICOM image). DICOMweb PACS Integration: DICOMweb WADO-RS (Web Access to DICOM Objects — Retrieve Service): retrieve DICOM images from any standards-compliant PACS. QIDO-RS (Query Instance): query for new studies in worklist. STOW-RS: store AI-annotated DICOM with triage annotations. Integration tested against: Sectra IDS7, Philips Vue PACS, Fujifilm Synapse — all three support DICOMweb. DICOM de-identification: patient identifiers stripped before AI inference (AI receives anonymised DICOM — re-identified for worklist display after inference). Radiologist Workflow Integration: Triage dashboard: worklist sorted by urgent finding probability (highest first). Radiologist overlay: AI confidence heatmap overlaid on DICOM viewer (indicating which region triggered the triage flag). One-click override: radiologist can downgrade AI urgent flag to routine with free-text reason (logged for post-market surveillance). Audit log: every AI triage decision, every radiologist override, and every report outcome — feeds post-market clinical follow-up (PMCF) study.