OncoView AI: Explainable Breast Imaging Triage Demo

Problem

Medical imaging models are often shown as a probability with no visual evidence, no operating-point choice, and no public way to test behaviour. That is a poor fit for a screening-style discussion where missing a malignant case is costly.

Solution

A research demo where a user picks a curated case or uploads an image, chooses balanced or high-recall mode, and inspects prediction, confidence, and heatmap overlays in one screen.

System Architecture

Primary data and decision flow

1. Study input DICOM or image uploads enter the same normalized inference path.
2. Ensemble inference Six EfficientNet checkpoints produce a combined risk score.
3. Operating point Balanced or high-recall mode applies the selected threshold.
4. Visual evidence Grad-CAM overlays and hotspots expose the regions behind the score.

1. Upload or gallery image is normalized and passed through a 6-checkpoint EfficientNet ensemble with optional test-time augmentation.
2. Balanced mode vs high-recall mode changes the operating point for more conservative screening behaviour.
3. Grad-CAM produces heatmap, overlay, and hotspot cues so the user can see which regions influenced the score.
4. Curated demo samples ship with the public service so the full private training set is not required at runtime.
5. Checkpoints live on Git LFS; the API runs as a systemd service behind Nginx.

What Dmitri Derevjanko built

Dmitri trained and packaged a 6-checkpoint EfficientNet ensemble, implemented FastAPI inference for DICOM/PNG/JPG, Grad-CAM overlays, two operating modes, a 20-case gallery, Git LFS for checkpoints, and 24/7 deploy with Nginx, systemd, and GitHub Actions. The GitHub repository documents the code.

Technologies

• Python
• FastAPI
• PyTorch
• EfficientNet
• Grad-CAM
• OpenCV
• Pydicom
• Git LFS

Result and Impact

A live explainable-imaging demo: ensemble inference, two modes, visual evidence, and a public GitHub repo, without shipping the full training dataset.

Limitations

• Research and portfolio demo only. Not a medical device, not CE-marked, not for diagnosis or screening in care.
• Performance on the demo gallery is not a clinical trial result.
• Users must not upload identifiable patient studies to a public demo.

Live demo and source

Live demo: https://oncoview.dmitriderevjanko.com/

GitHub: https://github.com/DmitriDerevjanko/cancer_detector

Portfolio case study: https://dmitriderevjanko.com/projects/oncoview-ai/

Other independent projects by Dmitri Derevjanko

• Law Agent
• Marketplace Intelligence
• OSINT Radar
• Global Risk Forecast
• SteelVision
• SmartCam
• FinanceLens Analyst
• Estonian Company Intelligence Hub
• Opportunity Intelligence Studio