Segmentation-Based Landmark Localization in Cerebral Magnetic Resonance Angiography Using Landmark Subsets
- Author:
Yura JEONG
1
;
Daehyun KWON
;
Se-On KIM
;
Ga-Hyeon KIM
;
Min-Seo PARK
;
Yoon-Chul KIM
Author Information
- Publication Type:Original Article
- From:Investigative Magnetic Resonance Imaging 2026;30(1):11-19
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:The accurate localization of bifurcation points in the brain near the circle of Willis is essential for labeling cerebral arteries and detecting potential aneurysms. Studies on the development of segmentation-based landmark localization methods for cerebral angiography are lacking. This study aimed to develop and validate a method for localizing anatomical landmarks using a three-dimensional (3D) encoder-decoder deep convolutional neural network architecture.
Materials and Methods:Time-of-flight magnetic resonance angiography (MRA) images of 224 subjects obtained from publicly available datasets were used. Ten anatomical landmark points were annotated, and four different landmark subset configurations were trained and validated. For each configuration, 3D U-Net-based models were developed using the MRA images and their corresponding annotated landmarks. The deep learning prediction results were evaluated in terms of landmark localization errors.
Results:Among the four configurations, the configuration with five landmark subsets produced the smallest landmark localization errors in the test dataset. Post-processing of the U-Net-predicted segmentation masks further reduced the mean localization errors across all landmark points for the configuration with five subsets.
Conclusion:The proposed landmark localization method effectively identified major anatomical landmarks around the circle of Willis and showed the potential to automate segmental analyses of intracranial arterial tortuosity.
