BMapEst: Estimation of Brain Tissue Probability Maps using a Differentiable MRI Simulator

[📖 Paper]

Introduction

In this paper, we introduce a novel framework leveraging the capabilities of MR-zero to estimate the tissue probability maps of digital brain phantoms representing the CSF, GM, and WM. Unlike supervised learning, our framework does not need a-priori training pairs of inputs and outputs and can further estimate tissue probability maps for any arbitrary set of MRI sequences (e.g., T₁-only, T₁+T₂,T₂+T₂^*+GRE, etc.) at any arbitrary echo times. Our contributions are:

1. We demonstrate the first method of estimating brain tissue probability maps using a differentiable MRI simulator that conducts forward inference to generate a T₁/T₂-weighted image by backpropagating a loss function to the brain tissue probability maps. Our approach is versatile, applicable to many different MRI sequences, and does not require learnable parameters.

2. We overcome the ill-posedness of probability maps estimation by using the inductive bias of the simulator and multiple T₁/T₂ contrasts.

3. We validate our approach on BrainWeb’s 20 subjects with the popular Fast Low Angle Shot (FLASH) sequence variants and obtain state-of-the-art results compared to supervised deep learning and clustering methods.

Installation

Python >= 3.10.12

git clone https://github.com/BioMedAI-UCSC/BMapEst.git
cd BMapEst
python3 -m venv bmapest
source bmapest/bin/activate
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

Due to memory constraints, estimation of tissue probability maps can only happen in CPU for this release.

Setup the BrainWeb dataset

Refer to the dataset.md file to obtain the BrainWeb slices.

Estimation

You can find the commands for all the experiments that are described in the paper in the subjects directory. Run all the experiments for a subject as follows:

bash subjects/subject04.sh

Cite BMapEst

If you find this code repository useful, please use the following BibTeX entry for citation.

@article{gupta2024bmapopt,
  title={BMapOpt: Optimization of Brain Tissue Probability Maps using a Differentiable MRI Simulator},
  author={Gupta, Utkarsh and Nikolakakis, Emmanouil and Zaiss, Moritz and Marinescu, Razvan},
  journal={arXiv preprint arXiv:2404.14739},
  year={2024}
}