DISINR

Code for Structure-Contrast Disentangled INRs for Accelerated Multi-contrast MRI Reconstruction, to be presented at MICCAI 2026 (link to be added when available).

Requirements

Install the Python dependencies with:

pip install -r requirements.txt

The current environment uses:

• torch
• sigpy
• h5py
• matplotlib
• scikit-image
• tqdm

Data

We use the publicly available M4RAW dataset. It contains multi-contrast brain MRI scans collected on a 0.3T custom scanner. All subject datasets should contain three contrasts: T1, T2, FLAIR, each with dimensions (18, 4, 256, 256) (Slices, Coils, Rows, Columns). Note that the M4RAW dataset contains multiple repetitions, which can be averaged for higher baseline SNR.

Demo

We have provided a demo notebook (demo.ipynb) that walks through the following:

1. Loading the M4RAW data
2. Instantiating a DISINR model
3. Setting up training environment
4. Reconstructing a slice from the loaded example subject
5. Visualizing reconstruction results and computing performance metrics

If you want to use the model directly in your own code, import DISINR from model.py and the helper utilities from utils.py.

Project Structure

.
├── demo.ipynb              # Notebook demonstrating how to use DISINR model for reconstructing a protocol
├── model.py                # DISINR model definition
├── utils.py                # Helper functions for making coordinate grids, joint TV loss, and generating sampling masks
└── requirements.txt        # List of required Python packages

License

See LICENSE for licensing details.