Medical Liver Segmentation ToolKit

Implemented Architectures

The following network architectures have been implemented: - UNet - VNet - ResUNet - UNet++ - Attention-UNet - UNETR - SwinUNETR - MedFormer - SegFormer

Results

| Architectures | DSC | NSD | MASD | HD | RVD | |------------------|----------|----------|----------|----------|----------| | UNet | 0.955 | 0.903 | 1.326 | 48 | -0.036 | | VNet | 0.941 | 0.860 | 3.483 | 148 | -0.052 | | ResUNet | 0.956 | 0.902 | 0.890 | 35 | -0.046 | | UNet++ | 0.951 | 0.888 | 1.218 | 37 | -0.009 | | Attention-UNet | 0.959| 0.907 | 0.845 | 37 | -0.037 | | UNETR | 0.952 | 0.890 | 1.263 | 36 | 0.006| | SwinUNETR | 0.956 | 0.898 | 0.837| 43 | -0.006| | MedFormer | 0.957 | 0.917| 1.251 | 55 | -0.021 | | SegFormer | 0.946 | 0.878 | 1.482 | 40 | -0.043 |

Quickstart

⚠️ Warning This project is under development

Installation

💡 Software installation Download Medical Liver Segmentation ToolKit

bash (.venv) $ git clone https://github.com/marqinhos/MedicalLiverSegmentationToolkit.git

bash (.venv) $ cd MedicalLiverSegmentationToolkit/

bash (.venv) $ pip install -r requirements.txt

Usage

❗ Important To introduce new network architectures, follow these steps:

1. Enter the architecture in model/dim3/{your_architecture}.py.
2. Add the network import in model/utils.py.
3. Add the network training configuration in config/{database}/{your_architecture}_3d.yaml.

Training Network

bash (.venv) $ python3 train.py --model {network_name} --max_epochs {num_max_epochs}

📝 Note See the Train Module documentation for more info on parameters.

💡 Tip To train more than 1 network, use train_sequential.py, more info.

Test Network

bash (.venv) $ python3 train.py --model {network_name} --version {training_version}

Predict Network

bash (.venv) $ python3 train.py --model {network_name} --version {training_version}

💡 Tip To predict more than 1 network, use predict_sequential.py, more info.

Evaluation of Trained Models

Performance Measures

📝 Note The performance measures computed are:

1. Dice Similarity Coefficient (DSC).
2. Normalize Surface Distance (NSD).
3. Mean Average Surface Distance (MASD).
4. Hausdorff Distance (HD).
5. Relative Volume Difference (RVD).

bash (.venv) $ python3 metrics_sequential.py

📝 Note Generate a JSON file with networks as shown below (more info): ```python :lineno-start: 20 :emphasize-lines: 10

models3d = [ 'attentionunet', 'medformer', 'resunet', 'swinunetr', 'unet++', 'unetr', 'vnet', 'segformer', '{namearchitecture}', ] ```

Complexity Measures

bash (.venv) $ python3 calculate_features_networks.py

📝 Note The profiler calculates:

1. Number of params.
2. Floating point operations per second (Flops).
3. Memory usage in an inference.
4. Layer sizes.

more info.

License

Medical Liver Segmentation ToolKit is available under next license:

• AGPL-3.0 License: See LICENSE file for details. ## Author: (c) 2024 (Marcos Fernández)