Model Card: CoronarySAM2 - Fine-tuned SAM2 for Coronary Artery Segmentation

Model Details

Model Description

CoronarySAM2 is a collection of fine-tuned Segment Anything Model 2 (SAM2) variants specifically optimized for coronary artery segmentation in X-ray angiography images. The models use point-based prompting to enable interactive and precise segmentation of coronary arteries from medical imaging data.

Developed by: Research Team
Model Type: Computer Vision - Image Segmentation
Base Architecture: SAM2 (Segment Anything Model 2) with Hiera backbone
Language(s): Python
License: [Specify License]
Fine-tuned from: facebook/segment-anything-2

Model Variants

Four model variants are available, offering different trade-offs between speed and accuracy:

Model	Parameters	Checkpoint	Speed	Accuracy	Use Case
SAM2 Hiera Tiny	~38M	`sam2_t/best_model.pt`	⚡⚡⚡ Fast	⭐⭐⭐ Good	Quick experiments, real-time feedback
SAM2 Hiera Small	~46M	`sam2_s/checkpoint_epoch_70.pt`	⚡⚡ Medium	⭐⭐⭐⭐ Very Good	Balanced performance, general use
SAM2 Hiera Base Plus	~80M	`sam2_b+/best_model.pt`	⚡ Slower	⭐⭐⭐⭐⭐ Excellent	High-quality results, clinical evaluation
SAM2 Hiera Large	~224M	`sam2_l/final_model.pt`	⚡ Slowest	⭐⭐⭐⭐⭐ Best	Maximum accuracy, research purposes

Model Architecture

The models follow the SAM2 architecture with the following components:

Image Encoder: Hiera hierarchical vision transformer backbone
Prompt Encoder: Encodes point prompts (positive/negative) as spatial embeddings
Mask Decoder: Transformer-based decoder that generates high-quality segmentation masks
Preprocessing Pipeline:
- X-ray image normalization using Gaussian blur
- CLAHE (Contrast Limited Adaptive Histogram Equalization) for vessel enhancement
- Fixed resolution resizing to 1024×1024 pixels

Intended Use

Primary Use Cases

Interactive Coronary Artery Segmentation: Point-based annotation for precise artery delineation
Medical Image Analysis: Automated assistance for cardiologists and radiologists
Research: Computer-aided diagnosis and treatment planning research
Educational Purposes: Training and demonstration of medical image segmentation

Out-of-Scope Use

❌ Clinical diagnosis without expert oversight
❌ Automated treatment decisions
❌ Real-time interventional guidance without validation
❌ Non-coronary vessel segmentation (not trained for this task)
❌ Modalities other than X-ray angiography (CT, MRI, etc.)

Training Data

Dataset

The models were fine-tuned on coronary X-ray angiography images with annotations for coronary artery structures.

Training Specifications:

Modality: X-ray Angiography
Target: Coronary Arteries
Annotation Type: Binary segmentation masks
Resolution: Images resized to 1024×1024 for training

Preprocessing

All training images underwent the following preprocessing pipeline:

Normalization: Gaussian blur-based intensity normalization
CLAHE Enhancement: Adaptive histogram equalization (clip limit: 2.0, tile grid: 8×8)
Resizing: Fixed 1024×1024 resolution
Format: RGB format (grayscale images converted to RGB)

Evaluation

Metrics

The models should be evaluated using the following metrics:

Dice Coefficient: Measures overlap between predicted and ground truth masks
IoU (Intersection over Union): Pixel-wise accuracy metric
Precision & Recall: For detecting true vessel pixels
Hausdorff Distance: Measures boundary accuracy
Inference Time: Speed benchmarks on various hardware

Performance Considerations

Point Prompt Quality: Model performance heavily depends on the quality and number of point prompts
Image Quality: Better results with high-contrast angiography images
Vessel Complexity: Performance may vary with vessel overlap and bifurcations
Model Selection: Larger models generally provide better accuracy but slower inference

How to Use

Installation

# Create conda environment
conda create -n sam2_FT_env python=3.10.0 -y
conda activate sam2_FT_env

# Install SAM2
git clone https://github.com/facebookresearch/segment-anything-2.git
cd segment-anything-2
pip install -e .
cd ..

# Install dependencies
pip install gradio opencv-python-headless torch torchvision torchaudio

Basic Usage

import torch
import numpy as np
from sam2.build_sam import build_sam2
from sam2.sam2_image_predictor import SAM2ImagePredictor

# Load model
checkpoint_path = "ft_models/sam2_s/checkpoint_epoch_70.pt"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

checkpoint = torch.load(checkpoint_path, map_location=device)
model_cfg = checkpoint['model_cfg']
sam2_model = build_sam2(model_cfg, checkpoint_path=None, device=device)

# Load state dict
state_dict = checkpoint['model_state_dict']
new_state_dict = {k[7:] if k.startswith('module.') else k: v 
                  for k, v in state_dict.items()}
sam2_model.load_state_dict(new_state_dict)
sam2_model.eval()

# Create predictor
predictor = SAM2ImagePredictor(sam2_model)

# Set image (preprocessed, 1024x1024, RGB, uint8)
predictor.set_image(preprocessed_image)

# Add point prompts
point_coords = np.array([[512, 300], [520, 310]])  # x, y coordinates
point_labels = np.array([1, 1])  # 1 = positive, 0 = negative

# Predict
masks, scores, logits = predictor.predict(
    point_coords=point_coords,
    point_labels=point_labels,
    multimask_output=True
)

Interactive Application

Launch the Gradio interface:

python app.py

Access at http://127.0.0.1:7860

Limitations

Technical Limitations

Fixed Input Size: Models expect 1024×1024 input (automatic resizing may affect small vessels)
Memory Requirements: Large model requires significant GPU memory (~8GB VRAM recommended)
Point Dependency: Requires manual point prompts; not fully automatic
Single Modality: Optimized only for X-ray angiography

Medical Limitations

Not FDA Approved: Not cleared for clinical diagnostic use
Requires Expert Review: All outputs must be validated by qualified professionals
Variability: Performance may vary across different imaging protocols and equipment
Edge Cases: May struggle with severe vessel overlap, calcifications, or poor image quality

Known Issues

High-contrast regions may cause over-segmentation
Thin vessel branches may be missed without precise point placement
Performance degradation on low-quality or motion-blurred images

Ethical Considerations

Medical AI Responsibility

Human Oversight Required: This tool is designed to assist, not replace, medical professionals
No Autonomous Decisions: Should never be used for automated clinical decisions
Training Data Bias: Model performance may reflect biases present in training data
Privacy: Ensure patient data is handled according to HIPAA/GDPR regulations

Fairness & Bias

Model performance across different patient demographics should be validated
Imaging equipment and protocols may affect performance
Consider potential biases in training dataset composition

Transparency

Model predictions should be explainable to medical professionals
Segmentation boundaries should be reviewable and editable
Point prompt influence on outputs should be clear to users

Citation

Base Model (SAM2)

@article{ravi2024sam2,
  title={SAM 2: Segment Anything in Images and Videos},
  author={Ravi, Nikhila and Gabeur, Valentin and Hu, Yuan-Ting and Hu, Ronghang and Ryali, Chaitanya and Ma, Tengyu and Khedr, Haitham and R{\"a}dle, Roman and Rolland, Chloe and Gustafson, Laura and others},
  journal={arXiv preprint arXiv:2408.00714},
  year={2024}
}

This Work

If you use CoronarySAM2 in your research, please cite:

@software{coronarysam2_2025,
  title={CoronarySAM2: Fine-tuned SAM2 for Coronary Artery Segmentation},
  author={[Your Name/Team]},
  year={2025},
  url={[Repository URL]}
}

Model Card Authors

[Primary Author Names]
Last Updated: November 2025

Contact

For questions, issues, or collaboration inquiries:

GitHub Issues: [Repository URL]/issues
Email: [Contact Email]

Disclaimer

⚠️ IMPORTANT MEDICAL DISCLAIMER ⚠️

This software is provided for research and educational purposes only. It is not intended for clinical use, medical diagnosis, or treatment planning. The models have not been validated for clinical deployment and are not FDA-approved or CE-marked medical devices.

Always consult qualified healthcare professionals for medical image interpretation and clinical decisions. The developers assume no liability for any clinical use or consequences resulting from the use of this software.

Additional Resources

Version: 1.0
Last Updated: November 18, 2025
Status: Research/Development

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for astroanand/CoronarySAM2

Base model

facebook/sam2-hiera-base-plus

Finetuned

(1)

this model

astroanand
/

CoronarySAM2