SAM-MedUI

Interactive Deep Learning for Myocardial Scar Segmentation Using Cardiovascular Magnetic Resonance

SAM-MedUI is a clinician-friendly interactive tool for myocardial scar segmentation from Late Gadolinium Enhancement Cardiac MRI (LGE-CMR). It combines a fine-tuned Segment Anything Model (SAM) with YOLO-based automatic detection and an intuitive GUI—enabling fast, accurate, and reproducible scar quantification without requiring any coding knowledge.

Features • Installation • Quick Start • Usage • Architecture • Citation

Screenshot

Key Highlights

Feature	Description
Clinical-Grade Interface	Built specifically for clinicians with intuitive controls and real-time feedback
Multiple Input Formats	Native support for DICOM, NIfTI (3D volumes), JPEG, PNG, and BMP
Flexible Prompting	Point-based, bounding box, and automatic YOLO detection
Real-time Refinement	Morphological operations, confidence adjustment, and undo
Quantitative Analysis	Automatic pixel mass calculations using DICOM/NIfTI metadata
Runs on CPU	GUI inference works on any laptop—no GPU required

Features

Interactive Segmentation

Point Prompts: Left-click to add foreground points (green)
Bounding Box Prompts: Click and drag to define regions of interest with adjustable handles
Auto-Detection: YOLO-based automatic cardiac region detection reduces manual prompting

Medical Imaging Support

DICOM: Full metadata extraction (patient ID, pixel spacing, slice thickness)
NIfTI: 3D volume support with automatic slice extraction and navigation
Standard Formats: JPEG, PNG, BMP for preprocessed images

Real-time Visualization

Mask Overlay: Alpha-blended segmentation visualization
Gamma Correction: Adjustable contrast (0.2–1.7) for enhanced visibility
Zoom & Pan: 0.5x to 5.0x magnification with smooth navigation

Mask Refinement

Morphological Operations: Expand/shrink masks with configurable iterations
Confidence Threshold: Dynamic adjustment (0.3–0.99) with real-time preview
Undo: Up to 10 levels of operation history

Batch Processing & Export

Thumbnail Gallery: Patient-centric navigation with multi-slice support
Batch Save: Export all masks with a single click
CSV Export: Quantitative results including patient ID, slice, and scar mass
Prompt Storage: JSON-based prompt saving for reproducibility

Installation

Prerequisites

Python 3.9 or higher
pip package manager
Git

Step 1: Clone the Repository

git clone https://github.com/Danialmoa/SAM-MedUI.git
cd SAM-MedUI

Step 2: Install Dependencies

pip install -r requirements.txt

Step 3: Run

Model weights are downloaded automatically on first launch from 🤗 Hugging Face.

cd GUI
python main.py

To download weights manually instead:

pip install huggingface_hub
huggingface-cli download AidaAIDL/SAM_MEDUI --local-dir checkpoints/

File	Description
`best_model.pth`	Fine-tuned SAM for cardiac scar segmentation
`yolo_best.pt`	YOLO detection model for automatic ROI detection

System Requirements

Component	Minimum	Recommended
RAM	8 GB	16 GB
GPU (Training)	CUDA-capable GPU	NVIDIA GPU with 8+ GB VRAM
GPU (Inference)	Not required	Optional for faster inference
Storage	2 GB for models	5+ GB with datasets

Quick Start

cd GUI
python main.py

Demo

Usage

Basic Workflow

Load Images → Click Load Folder or Load Files to import medical images
Add Prompts → Click for points, drag for bounding boxes, or use Auto-Detect
Generate → Click Generate Segmentation to create the mask
Refine → Adjust threshold or use Expand/Shrink for fine-tuning
Save → Export with Save Mask, Save All Masks, or Export Results

Keyboard Shortcuts

Shortcut	Action
`←` / `→`	Navigate between images
`Ctrl+Z`	Undo last operation
`Ctrl++` / `Ctrl+-`	Zoom in / out
`Ctrl+0`	Reset zoom
`Ctrl+Arrow keys`	Pan view
Hold `Z`	Temporarily hide mask & prompts

Prompting Modes

Mode	How to Use	Best For
Point (Foreground)	Left-click on target region	Precise selection of scar tissue
Bounding Box	Click and drag rectangle	Defining region of interest
Auto-Detect	Click the Auto-Detect button	Quick initial detection

Export Options

Save Mask: Export current segmentation as PNG
Save All Masks: Batch export all processed images
Export Results: Generate CSV with quantitative metrics:
- Patient ID
- Image/slice name
- Scar mass (calculated from pixel spacing and slice thickness)

Architecture

Project Structure

SAM-MedUI/
├── GUI/                              # Main Application
│   ├── main.py                       # GUI entry point and main window
│   ├── model_handler.py              # SAM & YOLO inference logic
│   ├── canvas_view.py                # Image display and annotation
│   ├── thumbnail_gallery.py          # Patient navigation and thumbnails
│   └── download_weights.py           # Auto-download weights from HuggingFace
│
├── SAM_finetune/                     # Training Pipeline
│   ├── models/
│   │   ├── sam_model.py              # SAM wrapper with fine-tuning support
│   │   ├── dataset.py                # Medical imaging dataset loader
│   │   ├── loss.py                   # Combined loss function
│   │   └── prompt_generator.py       # Bounding box & point generation
│   │
│   ├── train/
│   │   └── trainer.py                # Training loop with W&B logging
│   │
│   └── utils/
│       ├── config.py                 # Configuration dataclasses
│       ├── logger_func.py            # Logging setup with rotation
│       ├── preprocessing.py          # Image preprocessing utilities
│       ├── z_score_norm.py           # Percentile normalization
│       └── visualize.py              # Visualization helpers
│
├── checkpoints/                      # Model weights (auto-downloaded)
├── logs/                             # Application logs
├── requirements.txt                  # Python dependencies
├── setup.py                          # Package setup
└── README.md

Segmentation Pipeline

┌─────────────┐    ┌──────────────┐    ┌─────────────┐    ┌──────────────┐
│ Load Image  │───▶│ Add Prompts  │───▶│ SAM Forward │───▶│ Apply Mask   │
│ (DICOM/NIfTI)    │ (Points/BBox) │    │   Pass      │    │  Threshold   │
└─────────────┘    └──────────────┘    └─────────────┘    └──────────────┘
                          │                                       │
                   ┌──────▼──────┐                         ┌──────▼──────┐
                   │ YOLO Auto-  │                         │ Morphological│
                   │  Detection  │                         │  Refinement  │
                   └─────────────┘                         └─────────────┘

Training Pipeline

The fine-tuning pipeline includes:

Medical-specific augmentations via TorchIO (elastic deformation, motion artifacts, bias field)
Combined loss function: Dice + BCE + Soft BCE + KL Divergence + Diversity Loss
Experiment tracking with Weights & Biases
Learning rate scheduling with Cosine Annealing

Training (Fine-tuning)

To fine-tune SAM on your own cardiac MRI dataset:

from SAM_finetune.utils.config import SAMFinetuneConfig, SAMDatasetConfig
from SAM_finetune.models.dataset import SAMDataset
from SAM_finetune.train.trainer import SAMTrainer

# Configure dataset
dataset_config = SAMDatasetConfig(
    dataset_path="path/to/dataset",
    point_prompt=True,
    box_prompt=True,
    number_of_prompts=2,
)

# Create dataset
train_dataset = SAMDataset(config=dataset_config)

# Configure training
train_config = SAMFinetuneConfig(
    sam_path="pretrained_models/sam_vit_b_01ec64.pth",
    learning_rate=1e-4,
    num_epochs=100,
    batch_size=4,
)

# Start training
trainer = SAMTrainer(config=train_config, train_dataset=train_dataset)
trainer.train()

Citation

If you use SAM-MedUI in your research, please cite:

@article{moafi2026sammedui,
  title={Interactive Deep Learning for Myocardial Scar Segmentation Using Cardiovascular Magnetic Resonance},
  author={Moafi, Aida and Moafi, Danial and Shergil, Simran and Mirkes, Evgeny M. and Adlam, David and Samani, Nilesh J. and McCann, Gerry P. and Ghazi, Mostafa Mehdipour and Arnold, J. Ranjit},
  journal={Journal of Cardiovascular Magnetic Resonance},
  year={2026},
  publisher={Elsevier},
  url={https://www.sciencedirect.com/science/article/pii/S1097664726000384}
}

Authors

Aida Moafi¹, Danial Moafi², Simran Shergil¹, Evgeny M. Mirkes³, David Adlam¹⁵, Nilesh J. Samani¹⁵, Gerry P. McCann¹⁵, Mostafa Mehdipour Ghazi⁴*, J. Ranjit Arnold¹*

* Joint senior authorship

Affiliations

¹ Department of Cardiovascular Sciences, University of Leicester, NIHR Leicester Biomedical Research Centre and BHF Centre of Research Excellence, Glenfield Hospital, Leicester, UK ² Department of Information Engineering and Mathematics, University of Siena, Siena, Italy ³ Department of Mathematics, University of Leicester, Leicester, UK ⁴ Pioneer Centre for AI, Department of Computer Science, University of Copenhagen, Copenhagen, Denmark ⁵ Centre for Digital Health and Precision Medicine, University of Leicester

Acknowledgments

We gratefully acknowledge the following projects:

Segment Anything (SAM) by Meta AI Research
Ultralytics YOLO for object detection
TorchIO for medical image augmentation

Contact

For questions or collaborations:

Aida Moafi am1392@leicester.ac.uk Danial Moafi d.moafi@student.unisi.it

License

This project is licensed under the MIT License. See the LICENSE file for details.

Made with care for the medical imaging community

FilesExpand file tree

README.md

Latest commit

History