ALSAnchorNet: Using Multimodal MRI for ALS Diagnosis

Quick Start

1. Run Training Directly

python train.py

2. Run with Script

python run_training.py

Dataset

The cohort comprises 266 participants recruited from Qilu Hospital, including 159 patients with ALS and 107 age-matched healthy controls (HC). All participants were scanned on a 3T MRI system with rs-fMRI, DTI, and sMRI acquired. The processed dataset can be obtained by contacting Shuangwu Liu from the Qilu Hospital of Shangdong University (https://www.nursing.sdu.edu.cn/info/1201/5211.htm).

Data Format

The processed multimodal MRI data need to be placed in the ./data/ directory with the following organization:

• ./data/alps/: Processed ALPS data
• ./data/chp/: Processed ChP data
• ./data/fc/: FC matrices of all cases
• ./data/sc/: SC matrices of all cases
• ./data/fwf/: Processed fwf data
• ./data/gbold/: Processed gBOLD data
• ./data/pvs/: Processed PVS data

Note: If any of these data files is missing, synthetic random data will be generated for the purpose of testing.

Features

Integrated CAM Interpretability

• Automatically computes importance of frontal regions and connections during training
• Saves interpretability results only at best accuracy
• No additional configuration required

Complete Training Pipeline

• 5-fold cross-validation
• Early stopping mechanism
• Learning rate scheduling
• Resume from checkpoint

Automatic Result Saving

• Training logs: ./logs/
• Model files: ./models/
• CAM results: ./logs/cam_results/

Output File Structure

/workspace/
├── train.py                  # Main training file (with CAM integrated)
├── als_anchor_net_models.py  # ALSAnchorNet model definition
├── config.py                 # Configuration file
├── data_loader.py            # Data loader
├── logger.py                 # Logger
├── run_training.py           # One-click run script
├── models/                   # Model save directory
├── logs/                     # Log and CAM results directory
└── results/                  # Results directory