⛑SSM-Net: Semi-supervised multi-task network for a joint lesion segmentation and classification from pancreatic EUS images

SSM-Net aims to address joint segmentation and classification for the diagnosis of pancreas diseases. Specifically, for unlabeled EUS images, we establish an SRM to learn semantic saliency information for augmenting unlabeled images. By doing so, we can learn feature representation on unlabeled data for pre-training an encoder network to benefit the subsequent lesion segmentation and classification on labeled data. Then, we devise a series of modules, including an MGAM, a CAB and a FSL to boost joint lesion segmentation and classification.

Contents

1. Requirements
2. Environment Setup
   • Linux System
3. Dataset Preparation
4. Model Training and Evaluation

Requirements

This software is compatible with a Linux operating system, specifically Ubuntu 20.04 (compatibility with other versions has not been tested), and requires Python 3.7. It necessitates 64GB of RAM and 1TB of disk storage. Performance benchmarks are based on an Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz and an NVIDIA GeForce GTX 3090 Ti GPU.

The following Python packages are required, which are also listed in requirements.txt:

numpy==1.19.2
torch==1.13.1+cu117
torchvision==0.14.1+cu117
Pillow==9.2.0

Linux System

Step 1: Download the Project

1. Open the terminal, or press Ctrl+Alt+F1 to access the command line interface.
2. Clone this repository to your home directory.

git clone https://github.com/Torchlight-ljj/SSM-Net

3. Navigate to the cloned repository's directory.

cd SSM-Net

Step 2: Prepare the Environment and Execute the Code

1. Install the required Python packages.

python3 -m pip install --user -r requirements.txt

Supported Image File Formats JPEG and PNG file formats are supported and have been tested.

Steps

Training the Encoder Using MoCo Method

1. Place the images in ./moco_arch/data/train/0.
2. Run the script main.py to start training.
3. Note: Ensure the parameters are set correctly before running the script.

Training the Supervised Multi-Task Network

1. Navigate to the ./Sup_Multi_arch directory.
2. Organize the original images in ./Sup_Multi_arch/data/ori/... according to their class labels.
3. Place the corresponding masks in ./Sup_Multi_arch/data/mask/... according to their class labels.
4. Create train.txt and val.txt files for training and validation data.
5. Run the script train.py to initiate the training process.

Interpretability

• The CAM.py script generates the Class Activation Map (CAM) for the model's classification branch.