Using deep learning to automatically segment cardiac magnetic resonance imaging.

Model weights and independent application files (which can be used directly by clicking ".exe" file) are also availabel at: https://pan.baidu.com/s/1BM9viKgzGoECovzjxbMgtg?pwd=4396

🫀 CorSeg-CineSAX

Automatic Segmentation of Short-Axis Cine Cardiac MRI

基于深度学习的心脏短轴位电影磁共振（Cine SAX）自动分割工具

English · 中文说明

---

Overview

CorSeg-CineSAX is a desktop application for fully automatic segmentation of short-axis cine cardiac MRI. It provides an intuitive PyQt6 GUI for model inference, real-time visualization, and anatomical post-processing — no coding required.

Segmentation Targets

Label	Color	Structure
1	🔴 Red	Left Ventricle Myocardium (LV Myo)
2	🔵 Blue	Left Ventricle Blood Pool (LV Cavity)
3	🟢 Green	Right Ventricle Blood Pool (RV)

Key Features

• MedNeXt-L backbone — modern ConvNeXt-based medical image segmentation network
• PyQt6 GUI with real-time overlay visualization and opacity control
• Batch & single-file inference modes
• Three-step anatomical post-processing pipeline with violation detection
• DICOM & NIfTI input support (.dcm, .nii, .nii.gz)
• GPU (CUDA) / CPU selectable compute device
• Bilingual — English (_en) and Chinese (_ch) versions included
• JSON statistics report for post-processing results

---

中文说明

CorSeg-CineSAX 是一款基于深度学习的心脏短轴位电影 MRI 全自动分割桌面工具。提供直观的 PyQt6 图形界面，支持模型推理、实时可视化和解剖学后处理，无需编写代码。

分割目标

标签	颜色	结构
1	🔴 红色	左心室心肌 (LV Myocardium)
2	🔵 蓝色	左心室血池 (LV Blood Pool)
3	🟢 绿色	右心室血池 (RV Blood Pool)

---

Installation

1. Clone the Repository

git clone https://github.com/RunhaoXu2003/CorSeg.git
cd CorSeg

2. Create Environment (Recommended)

conda create -n corseg python=3.10 -y
conda activate corseg

3. Install Dependencies

# PyTorch (choose the command matching your CUDA version)
# See: https://pytorch.org/get-started/locally/
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# Core dependencies
pip install monai>=1.3 nibabel numpy matplotlib PyQt6

# Optional but recommended
pip install scipy       # Required for anatomical post-processing
pip install pydicom     # Required for DICOM file support

Requirements Summary

Package	Version	Required	Purpose
Python	≥ 3.9	✅	Runtime
PyTorch	≥ 2.0	✅	Deep learning backend
MONAI	≥ 1.3	✅	MedNeXt-L architecture
nibabel	≥ 3.0	✅	NIfTI I/O
numpy	≥ 1.21	✅	Array operations
matplotlib	≥ 3.5	✅	Visualization
PyQt6	≥ 6.4	✅	GUI framework
scipy	≥ 1.7	⬜	Post-processing
pydicom	≥ 2.3	⬜	DICOM support

---

Model Weights

Download the pre-trained model weights and place them in a folder (e.g., weights/):

weights/
└── best_model.pth

Download Link: [https://pan.baidu.com/s/1BM9viKgzGoECovzjxbMgtg?pwd=4396] (You can also get it from releases)

The .pth file contains:

• model_state_dict — MedNeXt-L trained weights
• config — training configuration (spatial dims, image size, etc.)

If the checkpoint contains a config key, it will be used automatically. Otherwise, default settings are applied:

{
    "spatial_dims": 2,
    "in_channels": 1,
    "num_classes": 4,
    "mednext_variant": "L",
    "mednext_kernel": 5,
    "img_size": (224, 224)
}

---

Usage

Launch the Application

# English version
python CorSeg-CineSAX_en.py

# 中文版本
python CorSeg-CineSAX_ch.py

Step-by-Step

1. Select inference mode — Folder (batch) or single file
2. Choose compute device — GPU (CUDA) or CPU
3. Set paths:
   • Model Folder — directory containing .pth weight file
   • Input Path — image file or folder with cardiac MRI data
   • Output Path — directory for saving segmentation masks
4. Configure post-processing (optional) — enable/disable individual steps
5. Click "Start Segmentation" — results display in real time on the right panel
6. Browse results — use Previous/Next buttons, adjust overlay opacity, toggle between raw prediction and post-processed views

Supported Input Formats

Format	Extension	Notes
NIfTI	.nii, .nii.gz	Recommended
DICOM	.dcm	Requires pydicom

Output Structure

Without post-processing:

output/
├── ROI-image001.nii.gz
├── ROI-image002.nii.gz
└── ...

With post-processing enabled:

output/
├── Prediction/          # Raw model predictions
│   ├── ROI-image001.nii.gz
│   └── ...
├── Postprocess/         # Post-processed results
│   ├── ROI-image001.nii.gz
│   └── ...
└── postprocess_stats.json   # Statistics report

---

Anatomical Post-Processing Pipeline

The three-step post-processing pipeline enforces anatomical priors on the model's raw predictions:

Step 1 — Connected Component Constraint

Remove isolated fragments: keep only the largest connected region per label.

Anatomically, each cardiac structure (LV Myo, LV Cavity, RV) should be a single contiguous region on each slice.

Step 2 — Containment Constraint

Ensure the LV blood pool (label 2) is fully enclosed by LV myocardium (label 1).

Exposed LV cavity pixels adjacent to background or RV are iteratively reassigned to myocardium.

Step 3 — Gap Filling Constraint

Fill background holes enclosed by cardiac structures.

• Part A: Holes fully enclosed by cardiac tissue are filled with the dominant neighboring label.
• Part B: Narrow gaps between LV myocardium and RV at the septum are filled.

Statistics Report

When post-processing is enabled, a postprocess_stats.json file is generated:

{
  "total_files": 100,
  "aggregate": {
    "before": {
      "fragment": 12,
      "containment_violation": 8,
      "gap": 5
    },
    "after": {
      "fragment": 0,
      "containment_violation": 0,
      "gap": 0
    }
  },
  "per_file": [
    {
      "file": "image001.nii.gz",
      "pre": { "has_fragment": true, "has_containment_violation": false, "has_gap": false },
      "post": { "has_fragment": false, "has_containment_violation": false, "has_gap": false },
      "pixels_changed": { "step1": 42, "step2": 0, "step3": 0 }
    }
  ]
}

---

Project Structure

CorSeg-CineSAX/
├── CorSeg-CineSAX_en.py    # English GUI application
├── CorSeg-CineSAX_ch.py    # Chinese (中文) GUI application
├── weights/                 # Model weights directory
│   └── best_model.pth      # Pre-trained MedNeXt-L weights
├── README.md                # This file
├── LICENSE                  # MIT License
└── assets/                  # (Optional) Screenshots, figures
    └── screenshot.png

---

Technical Details

Model Architecture

MedNeXt-L is a large-kernel ConvNeXt-based architecture designed for medical image segmentation, available in MONAI. Key characteristics:

• Large receptive field via 5×5 depthwise convolutions
• Efficient modern ConvNeXt design pattern
• 2D variant for slice-wise cardiac MRI segmentation

Inference Pipeline

Input Image (2D)
    │
    ▼
Resize to 224×224 (bilinear)
    │
    ▼
Z-score Normalization (non-zero voxels)
    │
    ▼
MedNeXt-L Forward Pass (FP16 on GPU)
    │
    ▼
Argmax → Predicted Labels
    │
    ▼
Resize to Original Resolution (nearest)
    │
    ▼
[Optional] Anatomical Post-Processing
    │
    ▼
Save as NIfTI (.nii.gz)

---

FAQ

Q: I get "MedNeXt is not available" error

Ensure MONAI ≥ 1.3 is installed:

pip install monai>=1.3

MedNeXt was added in MONAI 1.3. Older versions do not include it.

Q: Post-processing options are greyed out

Install scipy:

pip install scipy

Post-processing requires scipy for connected component analysis and morphological operations.

Q: DICOM files are not recognized

Install pydicom:

pip install pydicom

Q: Can I use this for 3D volumes?

The current version performs 2D slice-wise inference. For 3D volumes stored as multi-slice NIfTI files, the tool automatically extracts a representative 2D slice. For full 3D volume segmentation, process each slice individually or modify the inference pipeline.

Q: What GPU memory is required?

MedNeXt-L with 224×224 input requires approximately 2–4 GB GPU memory. Any modern NVIDIA GPU with CUDA support should work.

---

Citation

If you use this tool in your research, please consider citing: MedArxiv:

Related Works

• MedNeXt: Roy, S., et al. "MedNeXt: Transformer-driven Scaling of ConvNets for Medical Image Segmentation." MICCAI 2023.
• MONAI: The MONAI Consortium. "MONAI: Medical Open Network for AI." https://monai.io/

---

License

This project is licensed under the MIT License.

---

Made with ❤️ for the cardiac imaging community

If you find this useful, please ⭐ star the repository!