Bakary Gibba
Back to selected work

Case study · 05

Medx AI

A HER2 breast-cancer decision-support system combining histopathology classification, Grad-CAM explanations, LLM recommendation drafts, and clinician approval.

Explore on GitHubRead documentation
Category
Explainable clinical decision support
Availability
Research prototype · Public case study
Medx AI product interface and system workflow

Problem + approach

Overview

HER2 assessment shapes breast-cancer treatment, but ambiguous staining, observer variability, and opaque AI predictions limit clinical confidence.

Technology stack

  • Python
  • PyTorch
  • scikit-learn
  • Grad-CAM
  • Django
  • React
  • Groq
  • Gemini

Conceptual overview

System architecture

  1. 01Histopathology upload
  2. 02Quality + shift checks
  3. 03HER2 ensemble classifier
  4. 04Grad-CAM explanation
  5. 05LLM recommendation draft
  6. 06Clinician review + report

This project-specific overview is based on the published case-study architecture. Internal implementation, credentials, and deployment details remain undisclosed.

Decisions + constraints

Engineering notes

01

Engineering decisions

  • Model and LLM outputs stay assistive drafts; clinicians keep final responsibility.
  • Blur, stain, distribution-shift, and entropy checks can reject or defer uncertain inputs.
  • Grad-CAM evidence sits beside each prediction rather than in a separate notebook.
02

Technical challenges

  • Communicating uncertainty to clinical users
  • Separating doctor, nurse, and admin responsibilities with full traceability
03

Results & impact

  • ~96% reported weighted accuracy across four HER2 classes
  • 154 tests passed
  • 4.5 / 5 average user-acceptance score
04

Lessons learned

  • Clinical AI is a workflow problem as much as a model problem: review boundaries matter.

Screenshots + diagrams

Project media

Medx AI prediction result with Grad-CAM heatmap
Prediction + explanation — class probabilities beside a Grad-CAM++ heatmap.
Clinician review of an LLM recommendation draft
Clinician review — LLM drafts are accepted or rejected, never auto-applied.
Medx AI doctor dashboard
Doctor dashboard — cases, pending reviews, and traceable records.

Availability + roadmap

Documentation

Repository & API documentation

View repositoryRead full documentation

Future improvements

  • External validation
  • Calibration studies
Next case studyRednote Automation