Segmentation of Cerebral Arteries on 7T TOF MRA Images Using 3D U-Net
Poster No:
1649
Submission Type:
Abstract Submission
Authors:
Yuanzhe Huang1, George Stetten1, Jiatai Li2, Tamer Ibrahim1, Shaolin Yang1, Marcelo Rocha1, Howard Aizenstein1, Minjie Wu1
Institutions:
1University of Pittsburgh, Pittsburgh, PA, 2Carnegie Mellon University, Pittsburgh, PA
First Author:
Co-Author(s):
Introduction:
Accurate assessment of cerebral vasculature is a critical step in diagnosing and understanding various neurological disorders, including those characterized by subtle vascular changes as seen in Alzheimer's disease (Zimmerman et al., 2021). High-resolution medical imaging techniques, such as Time-of-Flight (TOF) Magnetic Resonance Angiography (MRA), facilitate the visualization of vascular structures and the identification of abnormalities. Automated segmentation methods are essential for efficiently delineating these vascular networks, thereby assisting clinicians in formulating well-informed treatment strategies. In this study, we explored the performance of a three-dimensional (3D) U-Net architecture (Çiçek et al., 2016) for segmenting cerebral vascular structures in 7T TOF MRA images.
Methods:
This study included 48 cognitively normal older adults (16 males, 32 females; mean age: 71.45 years), who underwent whole-brain TOF MRA scans acquired on a 7T Siemens scanner at the University of Pittsburgh. The imaging parameters included 354 slices and a voxel size of 0.38 × 0.38 × 0.38 mm³, with a total acquisition time of approximately 12 minutes.
Initial preprocessing involved manual skull stripping to remove non-brain tissues from each TOF image. The resulting brain mask was then dilated with a 2-mm spherical kernel to preserve vessels located along the brain surface. Ground truth datasets were generated using previously developed automated vessel segmentation algorithms. Specifically, Vessel Mapper (Li et al., 2023) was employed to derive a preliminary cerebral arterial tree. Subsequently, the express IntraCranial Arteries Breakdown (eICAB) (Dumais et al., 2022), an open-source deep learning algorithm for Circle of Willis segmentation, was utilized to identify key vascular landmarks. Start traversing from Circle of Willis, we calculated the vector between voxels along the medial axis of each vessel. Using this vector, a plane perpendicular to the vessel axis was determined, and intensity-based thresholds were applied to "unskeletonize" the vessel, thereby creating a refined ground truth.
A 3D U-Net (Çiçek et al., 2016) was trained using a training dataset of 45 TOF images, with an additional 3 images reserved for testing. The Adam optimizer (Kingma and Ba, 2014) was employed for 100 epochs, using an Intersection-over-Union (IoU) (Rezatofighi et al. 2019) loss function and an initial learning rate of 2 × 10⁻⁴. All training images were cropped or zero-padded to a uniform size of 368 × 512 × 416 to ensure dimensional consistency. A breadth-first search graph algorithm was applied as a post-processing step to identify the largest connected vascular cluster and refine the segmentation results.
Initial preprocessing involved manual skull stripping to remove non-brain tissues from each TOF image. The resulting brain mask was then dilated with a 2-mm spherical kernel to preserve vessels located along the brain surface. Ground truth datasets were generated using previously developed automated vessel segmentation algorithms. Specifically, Vessel Mapper (Li et al., 2023) was employed to derive a preliminary cerebral arterial tree. Subsequently, the express IntraCranial Arteries Breakdown (eICAB) (Dumais et al., 2022), an open-source deep learning algorithm for Circle of Willis segmentation, was utilized to identify key vascular landmarks. Start traversing from Circle of Willis, we calculated the vector between voxels along the medial axis of each vessel. Using this vector, a plane perpendicular to the vessel axis was determined, and intensity-based thresholds were applied to "unskeletonize" the vessel, thereby creating a refined ground truth.
A 3D U-Net (Çiçek et al., 2016) was trained using a training dataset of 45 TOF images, with an additional 3 images reserved for testing. The Adam optimizer (Kingma and Ba, 2014) was employed for 100 epochs, using an Intersection-over-Union (IoU) (Rezatofighi et al. 2019) loss function and an initial learning rate of 2 × 10⁻⁴. All training images were cropped or zero-padded to a uniform size of 368 × 512 × 416 to ensure dimensional consistency. A breadth-first search graph algorithm was applied as a post-processing step to identify the largest connected vascular cluster and refine the segmentation results.
Results:
Convergence of the training loss became evident at approximately 1500 steps. After 100 epochs, the model achieved an IoU loss of 0.3673 (Figure 1) and an average training accuracy of 99.88% (Figure 1). Evaluation on three held-out test subjects yielded an average test accuracy of 99.73%, demonstrating that the 3D U-Net (Çiçek et al., 2016) is highly effective in segmenting cerebral vasculature from 7T TOF MRA data (Figure 2).
In addition to improved accuracy, the proposed method substantially reduced processing time. Traditional vessel segmentation approaches, such as Vessel Mapper (Li et al., 2023) running on an Intel Xeon Platinum 8352Y CPU, required approximately 45 minutes per TOF scan. In contrast, the 3D U-Net, accelerated by GPU computation, required only around 10 seconds per image, thereby offering a dramatic improvement in efficiency.
In addition to improved accuracy, the proposed method substantially reduced processing time. Traditional vessel segmentation approaches, such as Vessel Mapper (Li et al., 2023) running on an Intel Xeon Platinum 8352Y CPU, required approximately 45 minutes per TOF scan. In contrast, the 3D U-Net, accelerated by GPU computation, required only around 10 seconds per image, thereby offering a dramatic improvement in efficiency.
Conclusions:
The 3D U-Net (Çiçek et al., 2016) provides rapid, accurate cerebral vessel segmentation from 7T TOF MRA, accelerating clinical workflows. However, it struggles with very small vessels, potentially limiting subtle pathology detection. Refinements in architecture, training diversity, and specialized techniques remain crucial for improving sensitivity to fine vascular structures.
Lifespan Development:
Aging 2
Modeling and Analysis Methods:
Segmentation and Parcellation 1
Keywords:
Cerebral Blood Flow
Cerebrovascular Disease
Data analysis
Machine Learning
MR ANGIOGRAPHY
1|2Indicates the priority used for review
·Figure 1. Training Loss and Training Accuracy
·Figure 2. Segmentation Result of Test subject
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
Li, J., Stetten, G., Schweitzer, N., Ibrahim, T., Iordanova, B., Aizenstein, H., Wu, M., (2023). VesselMapper - A Robust Vessel Segmentation Algorithm for 3D Images,, Organization for Human Brain Mapping (OHBM), Montréal, Canada, 2023.
Shi, Z., Li, J., Huang, Y., Schweitzer, N., Iordanova, B., Ibrahim, T., Yang, S., Stetten, G., Aizenstein, H., Wu, M., (2024). Exploring age-related morphological changes in cerebral arteries: a 7T TOF MRA study. Human Brain Mapping Conference, Seoul, Korea. June 23 - 27, 2024.
Dumais, F., Caceres, M.P., Janelle, F., Kassem, S., Arès-Bruneau, N., Gutierrez, J., Bocti, C., Whittingstall, K., (2022). eICAB: A novel deep learning pipeline for Circle of Willis multiclass segmentation and analysis, NeuroImage, Volume 260, 119425, ISSN 1053-8119, https://doi.org/10.1016/j.neuroimage.2022.119425.
Çiçek, Ö., Abdulkadir, A., Lienkamp, S., Brox, T., Ronneberger, O., (2016). 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. https://arxiv.org/abs/1606.06650
Kingma, D.P., Ba, J., (2014). A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, Silvio. (2019). Generalized Intersection over Union: A Metric and Loss for Bounding Box Regression, Computer Vision and Pattern Recognition.