Enhancement of PVS analysis via T1w-based PVS segmentation: Comparison with T2-weighted Based method
Poster No:
2006
Submission Type:
Abstract Submission
Authors:
Junghwa Kang1, Na-Young Shin2, Yoonho Nam1
Institutions:
1Division of Biomedical Engineering, Hankuk University of Foreign Studies, Yongin, Korea, Republic of, 2Department of Radiology, Severance Hospital, Yonsei University College of Medicine, Seoul, Select a State or Province
First Author:
Junghwa Kang
Division of Biomedical Engineering, Hankuk University of Foreign Studies
Yongin, Korea, Republic of
Division of Biomedical Engineering, Hankuk University of Foreign Studies
Yongin, Korea, Republic of
Co-Author(s):
Na-Young Shin
Department of Radiology, Severance Hospital, Yonsei University College of Medicine
Seoul, Select a State or Province
Department of Radiology, Severance Hospital, Yonsei University College of Medicine
Seoul, Select a State or Province
Yoonho Nam
Division of Biomedical Engineering, Hankuk University of Foreign Studies
Yongin, Korea, Republic of
Division of Biomedical Engineering, Hankuk University of Foreign Studies
Yongin, Korea, Republic of
Introduction:
Perivascular spaces (PVS), also known as Virchow-Robin spaces, play a crucial role in enveloping brain blood vessels [1]. Typically, their assessment involves 3D T2-weighted (T2w) imaging or a combination of T1-weighted (T1w) and T2w imaging[2,3]. However, the acquisition of 3D T2w images for PVS evaluation can be challenging in clinical or open datasets. This study introduces a segmentation method utilizing only T1w images, overcoming the limitations associated with T2w availability. We propose a two-step algorithm to enhance visibility and create a PVS mask, comparing its outcomes with the conventional T2w-based approach and providing an alternative for PVS quantification.
Methods:
[Dataset and Preprocessing]
We utilized 3T 3D T1w and T2w images from the Human Connectome Project (HCP) dataset (HCP-Young Adult, 22-35 years) [4]. Model training involved 927 subjects, and 45 subjects were used for model testing. Additionally, external validation was performed on an independent dataset (N=18, young-adults) obtained from 3T Philips MRI. Enhanced T1w targets for training were generated through pixel-wise division of T1w by T2w (T1w/T2w) after confirming co-registration accuracy.
[Network]
Our T1w-based segmentation method comprises two deep learning models, as depicted in Figure 1. We employed 3D U-Net and SwinUNetR[5] as the synthetic and segmentation models, respectively.
Model updates involved multiple loss functions, combining L1 Loss, minimum intensity projection (mIP), and L1 Loss with PVS-weighted map for the enhancement step. To calculate mIP Loss, we randomly chose six consecutive slices from the complete volume and applied minimum intensity projection (mIP) to both the output and target images, subsequently utilizing L1 Loss. For the segmentation step, a combination of Dice and cross-entropy loss functions was used.
[Evaluation]
To assess the results, SwinUNetR was also trained using single-contrast (T2w or T1w) input. PVS segmentation volume, and the number of connected components were calculated for comparison. Pearson correlation coefficients were computed between model results to measure similarity.
We utilized 3T 3D T1w and T2w images from the Human Connectome Project (HCP) dataset (HCP-Young Adult, 22-35 years) [4]. Model training involved 927 subjects, and 45 subjects were used for model testing. Additionally, external validation was performed on an independent dataset (N=18, young-adults) obtained from 3T Philips MRI. Enhanced T1w targets for training were generated through pixel-wise division of T1w by T2w (T1w/T2w) after confirming co-registration accuracy.
[Network]
Our T1w-based segmentation method comprises two deep learning models, as depicted in Figure 1. We employed 3D U-Net and SwinUNetR[5] as the synthetic and segmentation models, respectively.
Model updates involved multiple loss functions, combining L1 Loss, minimum intensity projection (mIP), and L1 Loss with PVS-weighted map for the enhancement step. To calculate mIP Loss, we randomly chose six consecutive slices from the complete volume and applied minimum intensity projection (mIP) to both the output and target images, subsequently utilizing L1 Loss. For the segmentation step, a combination of Dice and cross-entropy loss functions was used.
[Evaluation]
To assess the results, SwinUNetR was also trained using single-contrast (T2w or T1w) input. PVS segmentation volume, and the number of connected components were calculated for comparison. Pearson correlation coefficients were computed between model results to measure similarity.
·Figure 1. The pipeline of our proposed method.
Results:
Figure 2 summarized our result. Figure 2-A shows the improved PVS visibility compared to the original T1w image, as highlighted in the yellow circle. Figure 2-B compiles calculated PVS volumes and numbers from the segmentation results. The proposed method exhibited improved correlation coefficients for PVS volume (0.91 to 0.95) and PVS count (0.92 to 0.93), aligning closely with T2w results in Figure 2-b-1. In the external test set, no significant differences were observed in volume, but an improvement was noted in PVS count (Figure 2-B-2).
·Figure 2. Summarized our result. (A) is a representative example of results from the enhancement step. (B) is the PVS number and volume in White Matter for the test subjects (N=45).
Conclusions:
This study introduces a PVS segmentation method utilizing only 3D T1w images. Results from the T1-based method demonstrated high similarity to T2w results for PVS volume and count in the brain. Our approach is expected to enhance the clinical value of PVS quantification from 3D T1w images.
Modeling and Analysis Methods:
Segmentation and Parcellation 1
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Anatomy and Functional Systems 2
Neuroanatomy Other
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
Machine Learning
Segmentation
White Matter
1|2Indicates the priority used for review
Provide references using author date format
[2] Y. Choi, Y. Nam, Y. Choi, J. Kim, J. Jang, K.J. Ahn, B. Kim, N. Shin, MRI‐visible dilated perivascular spaces in healthy young adults: A twin heritability study, Hum. Brain Mapp. 41 (2020) 5313-5324.
[3] T. Rashid, H. Liu, J.B. Ware, K. Li, J.R. Romero, E. Fadaee, I.M. Nasrallah, S. Hilal, R.N. Bryan, T.M. Hughes, Deep learning based detection of enlarged perivascular spaces on brain MRI, Neuroimage: Reports. 3 (2023) 100162.
[4] D.C. Van Essen, K. Ugurbil, E. Auerbach, D. Barch, T.E. Behrens, R. Bucholz, A. Chang, L. Chen, M. Corbetta, S.W. Curtiss, The Human Connectome Project: a data acquisition perspective, Neuroimage. 62 (2012) 2222-2231.
[5] A. Hatamizadeh, V. Nath, Y. Tang, D. Yang, H.R. Roth, D. Xu, Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images, (2021) 272-284.