CAHAL Module

cahal.py implements the pipeline to use CAHAL [CLINICALLY APPLICABLE RESOLUTION ENHANCEMENT FOR LOW-RESOLUTION MRI SCANS] from native samples (raw):

1. Denoising using ANTs (linux library) DenoiseImage.
2. Resampling to 1×1×1 mm isotropic voxels using ANTs/ANTsPy.
3. CNN inference (single PSNR-best NN-UNet model per resolution/anisotropy cluster) for the CAHAL processing step.

This repository contains the pipeline script and expects model and data folders next to it.

Requirements

• Python 3.8+
• Packages:
  • nibabel
  • numpy
  • torch
• DenoiseImage executable from ANTs must be placed next to cahal.py and be executable (chmod +x).
• CUDA is optional; the script will use CPU if no GPU is available.

Install (example):

pip install nibabel numpy antspyx torch torchio

Files & Folders

• cahal.py — main pipeline script.
• models_*_half/ — trained model files. To download the models go to https://zenodo.org/records/23032648 and download T1w and FLAIR pth files.

Usage

Run the module directly:

python cahal_module_test.py

Behavior:

• Files in data/ are processed in sequence.
• If a filename does not contain denoised_, script runs DenoiseImage and creates processed_data/denoised_<orig>.
• If filename does not contain resampled_, script resamples to 1mm and creates processed_data/resampled_<name>.
• After preprocessing the script loads the appropriate model and runs inference, saving results as processed_data/CAHAL_processed_<name>.

You can skip denoising/resampling by pre-naming files with denoised_ and/or resampled_.

Logging / Output

The script uses simple timestamped, leveled prints:

• INFO — general steps and timings
• WARN — recoverable issues (e.g., empty volume)
• ERROR — failures
• OK — successful major steps

Log lines include a timestamp and an indent parameter for nested messages.

Notes & Tips

• Ensure voxel metadata (header/affine) is consistent; the script attempts to fix common header issues and sets output header zooms to 1.0 mm.
• The script expects input orientation alignment and will align to RAS internally.

Troubleshooting

• "Denoised file not created": verify DenoiseImage presence and permissions; check subprocess STDOUT/STDERR (printed on error).
• Model loading errors: ensure model filenames and the models/ folder exist and are readable.
• Memory / CUDA errors: try CPU mode (no CUDA) or reduce batch/volume size.