Medical Imaging AI
AI-Powered DICOM Analysis for Radiology
Overview
An on-premise AI platform that ingests DICOM studies from a clinic's PACS, runs vision-transformer inference for multi-pathology detection across chest X-rays and CT scans, and generates structured radiology reports via a RAG pipeline — cutting average report turnaround from 45 minutes to under 8 minutes while keeping all PHI on-site.
The Challenge
A private radiology clinic faced a growing backlog of DICOM studies: critical findings such as pneumothorax and pulmonary embolism queued alongside routine scans with no automated triage. Radiologists spent 40–50 minutes composing narrative reports manually, and strict data-governance rules prohibited sending medical images or patient identifiers to any cloud inference API.
The Solution
Built a self-hosted DICOM ingestion pipeline using Orthanc PACS with a FastAPI service subscribing to study-complete events via DIMSE. Deployed a fine-tuned BioViL-T vision transformer on-premise for multi-pathology detection with per-finding confidence scores and heatmap overlays. A Qdrant-backed RAG layer indexed radiology textbooks and clinical guidelines, feeding Claude-powered structured report generation with FHIR R4-compatible output. Radiologists review AI pre-reads and heatmaps in an OHIF Viewer dashboard before signing off, maintaining clinical authority over every diagnosis.
Results
- ✓Report turnaround cut from 45 minutes to under 8 minutes — an 82% reduction
- ✓94% sensitivity for critical findings (pneumothorax, pulmonary embolism) on internal validation set
- ✓Radiologist review backlog reduced by 65% within 6 weeks of go-live
- ✓Zero PHI transmitted to external cloud services — fully on-premise inference and storage
- ✓FHIR R4-compatible structured reports integrated directly into the clinic's EHR system
- ✓5× increase in studies processed per radiologist per shift at peak load