AI-Powered Clinical Decision Support System

Clinical Context

MedVision Italia, a network of 12 private radiology clinics across Lombardy, was facing a crisis common to European healthcare: a 35% increase in imaging volume with a simultaneous 20% shortage of qualified radiologists. Reading backlogs stretched to 72 hours for non-urgent cases, raising patient safety concerns.

The Solution: Multi-Modal AI Diagnostic Pipeline

We built a CE-marked (Class IIa) clinical decision support system that assists radiologists by pre-screening chest X-rays and CT scans, flagging critical findings, and generating structured preliminary reports.

1. Computer Vision Engine

• Fine-tuned a Vision Transformer (ViT-L/14) on a curated dataset of 2.3M anonymized medical images.
• Multi-label classification for 14 thoracic pathologies (pneumonia, cardiomegaly, pleural effusion, etc.).
• GradCAM++ heatmaps overlay directly on the DICOM viewer to explain why the model flagged a region.

2. Structured Report Generation

• A Retrieval-Augmented Generation layer queries the latest clinical guidelines (ACR Appropriateness Criteria) to ensure report language aligns with evidence-based practice.
• Reports are generated in natural Italian medical terminology and automatically populate the RIS/PACS system via HL7 FHIR interoperability.

3. Clinical Workflow Integration

• Seamlessly embedded into the existing PACS workflow — radiologists see AI findings as a "second reader" overlay.
• Priority Queue: Cases flagged as critical (pneumothorax, large effusion) are automatically escalated to the top of the worklist.
• Full audit trail for regulatory compliance (MDR 2017/745).

4. Privacy & Compliance

• All data stays on-premise — the model runs on NVIDIA DGX Station within the hospital data center.
• Federated Learning: Model improvements are shared across clinics without raw patient data ever leaving the premises.
• GDPR-compliant anonymization pipeline using k-anonymity and differential privacy.

Technology Stack

• AI/ML: PyTorch, Hugging Face Transformers, MONAI
• Backend: FastAPI, Celery (task queue), PostgreSQL
• Medical: HL7 FHIR, DICOM, Orthanc PACS integration
• Infrastructure: NVIDIA DGX Station, Docker, on-premise Kubernetes
• Frontend: React, Cornerstone.js (DICOM viewer), Tailwind CSS

Clinical Outcomes (6 months post-deployment)

• AUC 0.94 across 14 pathology classes (peer-reviewed).
• Reading time reduced by 45% for routine studies.
• Critical finding escalation: average time from scan to radiologist review dropped from 4.2 hours to 8 minutes.
• Zero false negatives on critical pneumothorax cases (100% sensitivity at 97.3% specificity).