UNet-LaparoSeg is a deep learning model designed for localizing and segmenting endometriosis lesions in laparoscopic images. It features a custom UNet architecture enhanced with a ResNet34 encoder, ASPP, and attention gates. The system also provides Persian-language diagnostic reports and visualization overlays.
📌 Live Demo (currently offline)
The web demo is temporarily disabled due to infrastructure limitations.
Previous access point: http://endovis.alihaghighat.ir/
- ✅ UNet-based segmentation with pretrained ResNet34 encoder
- 🧠 ASPP and Attention Gate modules for enhanced lesion localization
- 🧪 Color mask-to-label conversion with pixel tolerance
- 📦 Smart class-balanced sampling based on lesion size distribution
- 🧰 Albumentations-based data augmentation (elastic, dropout, flips, etc.)
- 📊 Full evaluation support: Precision, Recall, Dice, IoU, and more
- 📷 Auto-generation of overlay PNGs and transparent masks
- 📄 Persian medical-style report generation per test image
- ⚡ REST API via FastAPI for live inference and deployment
Input → Predicted Mask → Overlay
Reports are auto-generated in Persian, detailing lesion type and anatomical location (e.g., Endo-Ovary, Endo-Uterus, etc.).
.
├── api/ # FastAPI-based backend
│ ├── app.py # Main FastAPI application
│ ├── model.py # Inference logic
│ ├── best.pth # Trained model weights
│ ├── output_single/ # Prediction outputs (masks, overlays)
│ ├── test_frames/ # Input frames for testing
│ └── *.png # Temporary prediction result images
│
├── Glenda_v1.5_classes/ # Dataset and training artifacts
│ ├── annots/ # Annotation masks
│ ├── frames/ # Raw video frames
│ ├── logs_unet/ # Training logs
│ ├── masks_numeric/ # Numerical masks for segmentation
│ ├── split_vis/ # Train/val/test split visualizations
│ ├── callesification.ipynb # Notebook for classification/analysis
│ ├── coco.json # COCO-style annotation file
│ ├── config.json # Configuration file for model or pipeline
│ ├── label_colors.html # Color legend for label visualization
│ ├── labels.txt # List of semantic class labels
│ ├── loss_curve.png # Training loss curve
│ ├── model.py # Possibly alternate model definition
│ ├── statistics_overall.csv # Evaluation metrics and results
│ ├── val_dice_curve.png # Validation Dice score over epochs
│ └── Screenshot*.png # Miscellaneous screenshots
| Metric | Value |
|---|---|
| Dice (macro) | 0.0666 |
| IoU (macro) | 0.0243 |
| Recall | 0.0385 |
| Sensitivity | 0.3359 |
| Accuracy | 0.7565 |
⚠️ These metrics reflect real-world challenges such as class imbalance, visually subtle lesions, and small or irregular lesion shapes – common in real-world laparoscopic imagery.🔍 Note: High accuracy may be misleading due to dominance of background pixels. Metrics like Dice, IoU, and Recall are more meaningful for medical segmentation tasks.
This project is part of the LAP MAP research initiative focused on automatic diagnosis of endometriosis from laparoscopic imagery.
- 🛠 Model architecture, pipeline engineering, deployment & reporting: Ali Haghighat
- 💡 Research scope, dataset definition, clinical insight:
- All Python code (modeling, inference, API, visualization): © 2025 Ali Haghighat
- Research design, dataset preparation, clinical scope: Intellectual Property of LAP MAP Team
❗ Do not use this repository for commercial or clinical deployment without written permission from the authors.
- ✅ Confidence-aware scoring in diagnostic reports
- 🔍 Explainability tools (e.g., Grad-CAM integration)
- ☁️ Deployment packaging (Docker, Hugging Face Spaces, Streamlit Cloud)
- 🌐 Multilingual report generation (English + Persian)