Overview
This project focuses on detecting tuberculosis (TB) using the Mantoux PPD skin test. We developed an AI-driven solution leveraging U-Net++ for semantic segmentation to measure the induration reaction on the skin. The primary challenge was the limited dataset, requiring innovative solutions to achieve high accuracy.
Key Features
- AI Model Development: Used U-Net++ with features like deep supervision, optimized loss functions, and custom activation functions.
- Data Challenges: Tackled the small dataset using advanced augmentation techniques with Albumentations and OpenCV.
- Annotation Tools: Annotated images using CVAT for precise segmentation masks.
- End-to-End Deployment: Deployed on AWS EC2 with a Django application for staff and patient login.
Technologies Used
- PyTorch
- nnU-Net
- CVAT
- Albumentations
- Django
- Bootstrap 5
- AWS EC2
- OpenCV
Challenges Overcome
- Limited dataset (< 50 images) required advanced augmentation and preprocessing techniques.
- Achieving precision in medical-grade segmentation through rigorous tuning.
- Building a secure and scalable deployment using AWS EC2 and Django.
Visuals
Staff & Patient Login Page
Model Output Example
Annotation with CVAT