Loading
Back to Projects
AI & Medical Imaging93% Model Accuracy
Lung Cancer Detection System
Diagnostic tool that uses a fine-tuned Vision Transformer to flag suspicious CT scans, served via a secure Next.js dashboard.
Full-Stack Engineer
2024
Model Accuracy93.4%
On held-out test set
Inference Time<600ms
After quantization
Data EncryptionAES-256
At rest
Overview
A bridge between research and clinical practice. I fine-tuned a Vision Transformer (ViT) on localized CT scan datasets and built the secure Next.js app that radiologists use to upload scans, run inference, and review results — with accessible form flows for medical data.
Key Features
ViT-based classification with heatmap visualization
Secure DICOM upload + processing pipeline
Patient record management in PostgreSQL
Dockerized microservices for model inference
What Changed for the User
- 93.4% accuracy on held-out CT scan test set
- Inference under 600ms per scan after quantization
- AES-256 encryption at rest for all medical data
Technical Challenges
Real-time inference on standard server hardware.
Resolution
Model quantization + a specialized FastAPI/Uvicorn worker — inference time dropped from 4s to under 600ms.Strict patient privacy on large medical image datasets.
Resolution
AES-256 encryption at rest plus an automated anonymization pipeline before any data reaches the training loop.