IEEE GCON-26 · acceptedMedical AI2025 - 2026Project Lead
HealthATM
Lung nodule screening on edge devices
- sensitivity
- 91.4%sensitivity
- Dice score
- 0.84Dice score
- ROC-AUC
- 0.91ROC-AUC
- per CT study
- 22.5sper CT study
The problem
Early lung cancer screening needs radiologists and connectivity that most rural and underserved settings simply do not have. Most deep learning solutions assume high-end GPUs, constant connectivity and a specialist on hand.
What I built
An asynchronous, modular pipeline that runs locally: CT preprocessing, nodule detection, 3D segmentation, malignancy risk estimation and structured reporting - with explainability built in, not bolted on.
Highlights
- Explainability through segmentation masks and Grad-CAM saliency maps
- Structured findings that feed clinician reports - and patient-friendly summaries
- Lightweight models so a kiosk-class machine can do the work
How it works
- 01CT intakeDICOM study received
- 02PreprocessingNormalise & resample
- 03Nodule detection
- 043D segmentationMask per nodule
- 05Malignancy risk
- 06Structured reportClinician + patient view