This project was conducted at the A. A. Martinos Center for Biomedical Imaging, Harvard Medical School, under the supervision of Prof. Bastien Guerin.
In mri-to-ct, magnetic resonance images of the skull are converted into porosity maps using an image-to-image deep learning translation framework. Porosity maps are typically computed from CT scans, which are not always available in clinical trials due to the risks associated with exposure to ionizing radiation. mri-to-ct provides an accurate, time-efficient way to extract these maps from MR images, resulting in less inconvenience for scientists and patients alike. We believe this tool will address a need in the neuromodulation community, where porosity maps are needed to predict the propagation of acoustic waves through the skull.
This project builds on previous work by Koh et al.. We propose an original training method, backpropagation in the mask, which restricts the learning task to skull voxels and is enabled by the skull segmentation tool SAMSEG. We also contributed by systematically exploring different DL structures and hyperparameters.
The deep learning framework relies on a Generative Adversarial Network (GAN). After training, we use the generator (GEN) for porosity map inference, with MR images as inputs. In the study, we trained models compatible with either 2D or 3D images.
Create a new Python environment. To retrieve packages, run:
pip install -r requirements.txt
git clone https://github.com/MatDagommer/mri-to-ct.git
This program requires the following 3D volumes for each subject: CT, pseudo-CT, MRI, and skull mask. CT and MRI are raw volumes that can be retrieved with your own acquisition routine. Pseudo-CT can be obtained using SPM8-method [1]. The skull mask should be obtained with Freesurfer's segmentation tool SAMSEG [2].
In the folder /data/raw_data, create a folder for each subject containing 4 NIfTI files:
CTlnT1_resliced.nii(CT)pCTlnT1_resliced.nii(pseudo-CT)skull_SAMSEG_resliced.nii(Skull Mask)T1_resliced.nii(T1-MRI)
To start default preprocessing of the data, run:
python preprocessing.py --dataset <nameOfDataset>
You can customize data preprocessing. For more information on the different options, run:
python preprocessing.py --help
The resulting preprocessed data is stored in /data/datasets/<nameOfDataset>.
Run the following line to train the default model:
python train.py --dataset <nameOfDataset> --name <nameOfModel>
In the folder /notebooks, you will find result acquisition notebooks from the paper. You can use them as examples to load and plot your own results.
[1] Izquierdo-Garcia, D., Hansen, A. E., Förster, S., Benoit, D., Schachoff, S., Fürst, S., ... & Catana, C. (2014). An SPM8-based approach for attenuation correction combining segmentation and nonrigid template formation: application to simultaneous PET/MR brain imaging. Journal of Nuclear Medicine, 55(11), 1825-1830.
[2] Puonti, O., Iglesias, J. E., & Van Leemput, K. (2016). Fast and sequence-adaptive whole-brain segmentation using parametric Bayesian modeling. NeuroImage, 143, 235-249.