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 GroundingDINOcd GroundingDINO
pip install -e .# Install Segment Anythingcd 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