Skip to content

Latest commit

 

History

History
95 lines (73 loc) · 6.49 KB

README.md

File metadata and controls

95 lines (73 loc) · 6.49 KB
Raw

Impact of Initialization on Intra-Subject Pediatric Brain MR Image Registration: A Comparative Analysis Between SyN ANTs and Deep Learning-Based Approaches

Objectives

The objective of this project is to use deep learning (DL) techniques for registering pediatric brain MRI scans and inspecting different initialization methods. DeepReg version 0.0.0 [1] is used to implement the unsupervised learning-based registration task. The only main modification is the removal of affine data augmentation in the automatic pre-processing steps available in their framework. It is a tensorflow based implementation DL toolbox with unsupervised and weakly-supervised algorithms. The U-Net architecture was used and the output is a dense displacement field (DDF). One can easily train different networks using configuration files. The config files used in this work are available in this repository. DeepReg's GitHub repository [2] is available for further consultations with scripts coded in Python.

Requirements

DeepReg github repository from march 1st 2021 (version 0.0.0) and all its dependencies have to be installed in order to train the DL model with the yaml files. Refer to DeepReg documentation for specifics around the variables in the configuration files. To install all dependencies related to running DeepReg, run the following line of code to create a deepreg environment:

conda env create -f env_deepreg.yml

Dataset

The Calgary Preschool dataset is employed with its T1-weighted MRI images. A total of 64 subjects is utilized to have at least two time-point images per subject. The selected 64 subjects are presented in PatientDict.txt with the first column being the subject and the second all image scanIDs.

Preprocessing

Initially, the images underwent N4 bias field correction before being processed through SynthSeg version 2.0 to obtain 18 brain regions of interest for validation purposes. Subsequently, the images were rescaled to a 1.5 mm isotropic resolution using FLIRT version 6.0 with the (-applyisoxfm option).

Procedure

The pair-based registration (with registration done on all possible pairs (434 pairs)) evaluated three types of initialization approaches after pre-processing steps:

  • Non previously registered intra-subject pairs (NoReg)
  • Rigidly registered via ANTs intra-subject pairs (RigidReg)
  • Rigid and affine registered via ANTs intra-subject pairs (RigidAffineReg)

All three distinct inputs were employed in both DL-based and SyN ANTs [3] registration schemes, as illustrated in the figure below. A comparative analysis was conducted, considering Dice score results, pre-registration and registration time, the number of negative Jacobian determinants, and the sum absolute of log Jacobian determinants to evaluate their respective performance.

As for the full pipeline, it is visible below:

The scripts folder contains multiple functions and bash scripts for all analyses that were conducted, including:

  • Training all intra-subject pairs
  • Evaluating the registration learning-based approaches on segmentations after warping
  • Jacobian determinant calculations
  • Time calculations ANTs commands used to pre-register the images are available in DataHandle.py where the ANTs version used is 2.3.4.dev172-gc801b.

Analyses

Graphs depicting Dice score results in relation to the age interval between pairs are included in the article for white matter, gray matter, and cerebrospinal spinal fluid. Animated graphs have been generated to facilitate a more detailed examination of Dice scores per age interval for each pair at a local level across all 18 segmented regions. These regions are calculated by averaging right and left hemispheres for every region, except for the brain-stem, 3rd ventricle, 4th ventricle, and CSF, which are considered as a whole in the initial 32 given labels. The provided graphs illustrate the results for NoReg, RigidReg, and RigidAffineReg, with ANTs SyN Reg in red, DL Reg in green, and the initial alignment in blue.

References

[1]DeepReg. Image Registration with Deep Learning. 2021. url: https://deepreg.readthedocs.io/en/latest/tutorial/registration.html.
[2]DeepReg. Medical image registration using deep learning. 2021. url: https://github.com/DeepRegNet/DeepReg.
[3]Avants BB, Epstein CL, Grossman M, Gee JC. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal. 2008 Feb;12(1):26-41. doi: 10.1016/j.media.2007.06.004. Epub 2007 Jun 23. PMID: 17659998; PMCID: PMC2276735.

Citing this work

If some of these implementations helped you, please don't hesitate to cite the followings:

  • Previous Related Work in WBIR 2022 Proceedings:
@INPROCEEDINGS{Dimitrijevic2022-ns,
  title     = "Deep {Learning-Based} Longitudinal Intra-subject Registration of
               Pediatric Brain {MR} Images",
  booktitle = "Biomedical Image Registration",
  author    = "Dimitrijevic, Andjela and Noblet, Vincent and De Leener,
               Benjamin",
  publisher = "Springer International Publishing",
  pages     = "206--210",
  year      =  2022
}
  • MELBA 2024 Special Issue on Image Registration:
@ARTICLE{Dimitrijevic2024-pb,
  title     = "Impact of initialization on intra-subject pediatric brain {MR}
               image registration: A comparative analysis between {SyN} {ANTs}
               and deep learning-based approaches",
  author    = "Dimitrijevic, Andjela and Noblet, Vincent and De Leener,
               Benjamin",
  journal   = "J. Mach. Learn. Biomed. Imaging",
  publisher = "Machine Learning for Biomedical Imaging",
  volume    =  2,
  number    = "Image Registration",
  pages     = "916--955",
  month     =  jun,
  year      =  2024,
  language  = "en"
}