SynPED

Overview

This project focuses on intelligent assisted diagnosis of precancerous lesions, supporting analysis of endoscopic images for:

• Classification: Identifying whether images contain precancerous lesions (0 - no lesion, 1 - precancerous lesion)
• Detection: Using GroundingDINO for object detection to locate lesion areas
• Segmentation: Using Segment Anything (SAM) for semantic segmentation to generate precise lesion region masks

Project Structure

SynPED/
├── src/                    # Core source code
│   ├── config/            # Model configuration files
│   ├── datasets/          # Dataset definitions
│   ├── models/            # Model implementations
│   │   └── modeling.py    # SynPED main model
│   └── utils/             # Utility functions
├── scripts/               # Evaluation and inference scripts
│   ├── inference.py      # Single image inference example
│   ├── eval_classification.py
│   ├── eval_detection.py
│   ├── eval_segmentation.py
│   └── eval_synped.py
├── GroundingDINO/        # GroundingDINO dependency
├── segment_anything/      # SAM dependency
├── clip/                  # CLIP dependency
└── wise-ft/              # Incremental fine-tuning related code

Installation

# Install GroundingDINO
cd GroundingDINO
pip install -e .

# Install Segment Anything
cd segment_anything
pip install -e .

Quick Start

Single Image Inference

Configure model paths in scripts/inference.py and run the script:

clip_path = "/path/to/finetuned_classifier.pt"
dino_config_file = "src/config/GroundingDINO_SwinT_OGC.py"
dino_path = "/path/to/dino_checkpoint.pth"
sam_path = "/path/to/medsam_model.pth"

Acknowledgments

• GroundingDINO
• Segment Anything
• CLIP