Skip to content
All work
Medical AI

CRC-MedAI

Real-time clinical decision support for colorectal cancer, with computer vision wired into a hospital workflow.

Year
2026
Role
Sole engineer & researcher (M.Sc. thesis)
Status
Research · IEEE CBMS 2026 target
CRC-MedAI: Clinical dashboard: live KPIs, population recurrence-risk trend and per-service health (Auth, GPU inference, prognosis, FHIR gateway).
Clinical dashboard: live KPIs, population recurrence-risk trend and per-service health (Auth, GPU inference, prognosis, FHIR gateway).

The problem

Colorectal cancer screening generates more imaging than clinicians can read quickly, and most research models never leave a notebook. The gap is not the model. It is everything around it: getting a prediction to a clinician, inside their workflow, fast enough to use live and explainable enough to trust.

The approach

I built a dual-branch system that handles detection and prognosis, then wrapped it in the engineering a hospital actually needs. The detection branch runs real-time procedure analysis with Grad-CAM heatmaps over the endoscopy stream, so a flagged polyp comes with the reason it was flagged. The prognosis branch predicts recurrence risk and exposes it through SHAP feature importance, decision paths and a what-if simulator, so a clinician can interrogate the score rather than take it on faith. End-to-end ML pipelines cover training, evaluation, inference and deployment under MLOps discipline, and FHIR-based APIs let it speak the language of clinical record systems. The backend is a set of containerized microservices, so the model, the API and the record integration scale and fail independently.

The result

A working real-time pipeline: video and patient data go in, and structured, explainable predictions come back through standard healthcare APIs. Inference latency, throughput and per-service health are tracked live in Grafana, so the model is observable in production instead of being a black box. Supervised at ESTIN with CHU de Béjaïa, and targeted at IEEE CBMS 2026.

Highlights

  • Dual-branch system: real-time detection plus recurrence prognosis
  • Explainability built in: Grad-CAM heatmaps for detection, SHAP and a what-if simulator for prognosis
  • FHIR interoperability so it plugs into clinical record systems
  • End-to-end MLOps pipeline, from training to deployment
  • Live inference monitoring (latency, throughput, health) in Prometheus and Grafana
  • Containerized microservice backend

Stack

  • PyTorch
  • Computer Vision
  • MLOps
  • FHIR
  • FastAPI
  • Docker
  • Microservices
  • Prometheus
  • Grafana

Inside the build

CRC-MedAI: Real-time procedure analysis: Grad-CAM heatmaps and live polyp detection with bounding boxes and confidence, straight off the endoscopy stream.
Real-time procedure analysis: Grad-CAM heatmaps and live polyp detection with bounding boxes and confidence, straight off the endoscopy stream.
CRC-MedAI: Recurrence Prognosis Engine with Explainability AI: SHAP global feature importance, decision paths and a what-if simulator.
Recurrence Prognosis Engine with Explainability AI: SHAP global feature importance, decision paths and a what-if simulator.
CRC-MedAI: Per-patient recurrence-risk prediction, run from the same review surface.
Per-patient recurrence-risk prediction, run from the same review surface.
Next project

RSM Suite