Motivation
Diabetic foot ulcers are a leading cause of preventable amputations, and early detection dramatically improves outcomes. Specialist screening doesn't scale to millions of patients — so the goal was an automated detector accurate enough for real clinical screening and deployable at population scale.
Architecture
Clinical images are annotated and augmented in Roboflow, then used to train a YOLO detector that localizes ulcers in new images for triage.
📷Clinical Images3,000+
→
🏷️Annotate + AugmentRoboflow
→
🧠YOLO TrainingPyTorch
→
🎯Ulcer Detection95% accuracy
From raw clinical photos to deployable triage detection.
What I built
- A deep-learning detector for diabetic foot ulcers, now adopted by the Tamil Nadu State Government.
- A YOLO-based detection model at 95% accuracy, with potential impact across 10M+ patients.
- Improved robustness 15% by annotating and augmenting 3,000+ images in Roboflow.
- Owned the full lifecycle: preprocessing, training, evaluation and deployment.
Results
95%Detection accuracy
+15%Robustness from augmentation
10M+Potential patient reach
GovtState-adopted
Challenges
- Limited, messy medical data: careful annotation and augmentation of 3,000+ images was essential to generalize.
- Clinical stakes: false negatives are costly, so evaluation focused on recall, not just headline accuracy.
- Deployability: the model had to run reliably outside a research notebook for government use.
Learnings
- Data quality beats model choice in medical CV — augmentation moved the needle more than architecture tweaks.
- In healthcare, the right metric (recall/sensitivity) matters more than raw accuracy.
- Real impact comes from shipping — a deployed 95% model beats a perfect notebook.