AnatGuidedCMBSeg

Official implementation of Anatomically-Guided Segmentation of Cerebral Microbleeds in T1-weighted and T2*-weighted MRI.

News

• (Jun. 2024) Our paper has been accepted to MICCAI 2024!

Citation

If you find this code useful in your research, please consider citing:

    @inproceedings{kwon2024AnatGuided,
	author={Kwon, Junmo and Seo, Sang Won and Park, Hyunjin},
	title={Anatomically-Guided Segmentation of Cerebral Microbleeds in T1-weighted and T2*-weighted MRI},
	booktitle={27th International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
	pages={24--33},
	doi={10.1007/978-3-031-72069-7_3},
	year={2024},
    }

Installation

1. Install Python 3.8 and PyTorch 1.11.0. We recommend to use pytorch/pytorch:1.11.0-cuda11.3-cudnn8-runtime
2. Install nnUNet-v1
3. Install SimpleITK
4. git clone https://github.com/junmokwon/AnatGuidedCMBSeg
5. Install ANTs and FreeSurfer including SynthStrip for MRI preprocessing.

T1-weighted MRI Preprocessing

1. Ensure that T1 and T2* MRI scans have the same orientation, resolution, and fields of view.
2. Ensure there are no oblique or orientation issues in T1-weighted MRI data.
3. Run Freesurfer recon-all

T2*-weighted MRI Preprocessing

1. Ensure that T1 and T2* MRI scans have the same orientation, resolution, and fields of view.
2. Ensure there are no oblique or orientation issues in T2*-weighted MRI data.
3. Run SynthStrip skull stripping mri_synthstrip
4. Run N4ITK N4BiasFieldCorrection
5. Perform rigid-body registration from T1 space to T2* space.

JHU-DTI Atlas Preprocessing

1. Download JHU-DTI Atlas
2. Prepare white-matter atlases including internal capsule and external capsule.
3. Perform rigid-body registration antsRegistrationSyN.sh -t r from MNI152 space to MNI305 (Talairach) space.
4. Transform IC and EC labels to Talairach space.

Proxy Label Generation

1. Generate lobar parcellation to obtain aparc.lobes.mgz and wmparc.lobes.mgz
2. Merge cerebral lobes from aparc.lobes.mgz and wmparc.lobes.mgz into lobar region.
3. Transform internal capsule and external capsule labels from MNI305 (Talairach) space to subject’s native T1 space using Talairach transform transforms/talairach.xfm
4. Merge deep white matter regions from aparc.lobes.mgz, wmparc.lobes.mgz, and JHU-DTI atlas into deep supratentorial region.
5. Merge brainstem and cerebellum from aparc.lobes.mgz into infratentorial region.

nn-UNet Training

1. Preprocess T1 and T2* MRI scans.
2. Generate target labels: lobar, deep supratentorial, infratentorial, and CMB labels.
3. Choose a task ID e.g., Task301_InHouse and Task302_VALDO2021
4. Run nnUNet_plan_and_preprocess -t 301 where 301 is Task ID.
5. Train a 3D full resolution nn-UNet with nnUNetTrainerV2_Loss_DiceTopK10 trainer.

Clinically-derived False Positive Reduction

1. Run nnUNet_predict to predict hard-thresholded segmentation masks for proxy and CMB labels.
2. For each connected component in CMB prediction masks, do the following:
3. Generate peripheral mask by performing logical XOR between dilated mask and predicted mask.
4. Calculate brain parenchyma ratio in peripheral mask using proxy prediction masks.
5. Discard the connected component if the ratio does not exceed 0.5