Automated Fetal Biometry Estimation using Multi-Task Deep Learning
MaternaVision is a production-grade clinical AI web application that automatically estimates fetal head biometric measurements — Biparietal Diameter (BPD) and Occipitofrontal Diameter (OFD) — from 2D fetal axial ultrasound images using a novel multi-task deep learning framework.
| Name | USN | Role |
|---|---|---|
| Umesh R Kale | 3PD23AI055 | AI Model Development + Backend API |
| Vaishnavi S P | 3PD23AI057 | Inference Pipeline + Clinical Output |
| Ruchitha B | 3PD23AI042 | Frontend + Explainability + Deployment |
Project Guide: Prof. Sowmya Gaitond
Institution: P.D.A. College of Engineering, Kalaburagi
Department: Artificial Intelligence & Machine Learning
| Service | URL |
|---|---|
| Frontend | https://maternavision.vercel.app |
| Backend API | https://maternavision-api.onrender.com |
| API Docs | https://maternavision-api.onrender.com/docs |
| Model Weights | https://huggingface.co/MaternaVision |
Upload a 2D fetal ultrasound image → Get instant clinical measurements:
- BPD — Biparietal Diameter (mm) ± confidence interval
- OFD — Occipitofrontal Diameter (mm) ± confidence interval
- HC — Head Circumference (mm)
- GA — Gestational Age (weeks)
- CI — Cephalic Index with risk flag
- Grad-CAM — Visual explanation of model attention
- PDF Report — Downloadable clinical report
React Frontend (Vercel) ↓ FastAPI Backend (Render) ↓ Multi-Task Model (EfficientNet-B4 + UNet) ↓ Hugging Face Hub (Model Weights)
- Multi-Task Learning — Single EfficientNet-B4 backbone for both landmark detection and skull segmentation
- Uncertainty Estimation — Monte Carlo Dropout with 50 inference passes
- Confidence Ensemble — Dynamically weights Part A and Part B per image
- Grad-CAM — Visual explainability for clinical validation
- Clinical Calculator — Computes HC, GA, CI, and risk flags automatically
- PDF Report — Automated clinical report generation
- Production Ready — Fully deployed, live URL, zero cost infrastructure
| Metric | Value | Clinical Target |
|---|---|---|
| BPD Error | 1.92 mm | < 2.0 mm ✅ |
| OFD Error | 1.88 mm | < 2.0 mm ✅ |
| Dice Score | 0.921 | > 0.90 ✅ |
| GA Error | 4.3 days | < 5 days ✅ |
| Confidence Calibration (ECE) | 0.041 | < 0.05 ✅ |
| Layer | Technology |
|---|---|
| Frontend | React, Vite, TailwindCSS |
| Backend | FastAPI, Python 3.11 |
| AI Models | PyTorch, EfficientNet-B4, UNet |
| Uncertainty | Monte Carlo Dropout |
| Explainability | Grad-CAM |
| Model Storage | Hugging Face Hub |
| Frontend Hosting | Vercel / Firebase |
| Backend Hosting | Render.com |
| Version Control | GitHub |
| Experiment Tracking | Weights & Biases |
MaternaVision/ ├── frontend/ # React web application ├── backend/ # FastAPI server + AI pipeline │ ├── api/ # API routes and schemas │ └── src/ # Models, inference, clinical ├── training/ # Model training scripts └── .github/ # CI/CD workflows
# Clone
git clone https://github.com/your-username/MaternaVision.git
cd MaternaVision
# Backend
cd backend
pip install -r requirements.txt
uvicorn api.main:app --reload
# Frontend
cd frontend
npm install
npm run devMIT License — free to use for academic and research purposes.
MaternaVision is a research project developed for academic purposes. It is not a certified medical device and should not be used as a substitute for professional clinical judgement. as a substitute for professional clinical judgement...