Print  Close

VesselBoost: Improving Segmentation of Small Vessels in Human Brain Magnetic Resonance Angiograms

Poster No:

2007 

Submission Type:

Abstract Submission 

Authors:

Marshall Xu1, Fernanda Ribeiro1, Siyu Liu1,2, Thomas Shaw3, Hendrik Mattern4,5,6, Soumick Chatterjee4,7,8, Oliver Speck4,5,6, Omer Faruk Gulban9,10, Grant Hartung11, Steffen Bollmann1,12, Jonathan Polimeni11,13,14, Markus Barth1,3, Saskia Bollmann1

Institutions:

1School of Electrical Engineering and Computer Science, The University of Queensland, Brisbane, Australia, 2Australian eHealth Research Centre, CSIRO, Brisbane, Australia, 3Centre for Advanced Imaging, The University of Queensland, Brisbane, Australia, 4Institute of Experimental Physics, Otto-von-Guericke-University, Magdeburg, Germany, 5German Center for Neurodegenerative Diseases, Magdeburg, Germany, 6Center for Behavioral Brain Sciences, Magdeburg, Germany, 7Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, 8Genomics Research Centre, Human Technopole, Milan, Italy, 9Faculty of Psychology and Neuroscience, Maastricht University, Maastricht, Netherlands, 10Brain Innovation, Maastricht,, Netherlands, 11Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, USA, 12Queensland Digital Health Centre, The University of Queensland, Brisbane, Australia, 13Department of Radiology, Harvard Medical School, Boston, United States, 14Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, United States

First Author:

Marshall Xu  
School of Electrical Engineering and Computer Science, The University of Queensland
Brisbane, Australia

Co-Author(s):

Fernanda Ribeiro  
School of Electrical Engineering and Computer Science, The University of Queensland
Brisbane, Australia
Siyu Liu  
School of Electrical Engineering and Computer Science, The University of Queensland|Australian eHealth Research Centre, CSIRO
Brisbane, Australia|Brisbane, Australia
Thomas Shaw, PhD  
Centre for Advanced Imaging, The University of Queensland
Brisbane, Australia
Hendrik Mattern  
Institute of Experimental Physics, Otto-von-Guericke-University|German Center for Neurodegenerative Diseases|Center for Behavioral Brain Sciences
Magdeburg, Germany|Magdeburg, Germany|Magdeburg, Germany
Soumick Chatterjee  
Institute of Experimental Physics, Otto-von-Guericke-University|Faculty of Computer Science, Otto von Guericke University|Genomics Research Centre, Human Technopole
Magdeburg, Germany|Magdeburg, Germany|Milan, Italy
Oliver Speck  
Institute of Experimental Physics, Otto-von-Guericke-University|German Center for Neurodegenerative Diseases|Center for Behavioral Brain Sciences
Magdeburg, Germany|Magdeburg, Germany|Magdeburg, Germany
Omer Faruk Gulban, Ph.D.  
Faculty of Psychology and Neuroscience, Maastricht University|Brain Innovation
Maastricht, Netherlands|Maastricht,, Netherlands
Grant Hartung  
Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital
Charlestown, USA
Steffen Bollmann  
School of Electrical Engineering and Computer Science, The University of Queensland|Queensland Digital Health Centre, The University of Queensland
Brisbane, Australia|Brisbane, Australia
Jonathan Polimeni  
Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital|Department of Radiology, Harvard Medical School|Division of Health Sciences and Technology, Massachusetts Institute of Technology
Charlestown, USA|Boston, United States|Cambridge, United States
Markus Barth  
School of Electrical Engineering and Computer Science, The University of Queensland|Centre for Advanced Imaging, The University of Queensland
Brisbane, Australia|Brisbane, Australia
Saskia Bollmann  
School of Electrical Engineering and Computer Science, The University of Queensland
Brisbane, Australia

Introduction:

The sophisticated pial arterial vascular network plays an important role in understanding cerebral blood flow dynamics and diagnosing cerebrovascular diseases. Starting by leveraging deep-learning models to extract accurate segmentations of the trivial vessels in human brains, we aim to quantitatively characterize the architecture of the pial arterial network to provide valuable insights for cerebral blood flow estimation models and fMRI signal simulations.

Methods:

Deep-learning methods for vessel segmentation:
VesselBoost offers two main avenues for detailed vasculature segmentation:
1. Test Time Adaptation (TTA): VesselBoost adapts a pre-trained model to new data, where an imperfect segmentation is available or generated with the pre-trained model. Pre-trained models were trained on 300 μm isotropic angiograms with manual or automatic labels.
2. Self-training Model: VesselBoost trains a 3D-UNet model from scratch using a single subject's imperfect segmentation for 'boosting' vessel details in the single subject data.
Both TTA and the Self-training Model incorporate an innovative data augmentation technique that utilizes the self-similarity between large and small vessels to expand training dataset.
Supporting Image: figure1_v2.png
   ·Figure 1: VesselBoost segmentation pipelines for TTA and Self-training Model
 

Results:

VesselBoost generated detailed segmentation across different data set with different resolutions. Figure 2 shows that both TTA and the Self-training Model can enhance the details and improve the continuity of the vessels in imperfect segmentations. Moreover, the TTA method exhibits strong generalization capabilities across varying resolutions, despite the pre-trained models being initially trained on lower-resolution data, it has a good performance on data with 150µm isotopic resolution.
Supporting Image: figure2_v3.png
   ·Figure 2: VesselBoost segmentation results from TTA and Self-training Model
 

Conclusions:

We developed a segmentation framework capable of extracting small vessels in magnetic resonance angiograms. This is an open-source project and currently available on https://github.com/KMarshallX/vessel_code. Future efforts are needed to firstly enhance the precision of vascular graphing, to extract quantitative measures and provide valuable input for blood flow simulations. Second future direction will be to increase VesselBoost's generalizability on other contrasts: T1, T2*, etc. In summary, VesselBoost demonstrates effective segmentation of small vessels, enhancing vascular continuity, while the TTA method successfully translates data from low to high resolution.

Modeling and Analysis Methods:

Methods Development 2
Segmentation and Parcellation 1

Neuroanatomy, Physiology, Metabolism and Neurotransmission:

Anatomy and Functional Systems

Keywords:

MR ANGIOGRAPHY
Other - Pial Arterial Vasculature; Deep-learning; Segmentation; Test-time Adaptation; Vascular Characteristics

1|2Indicates the priority used for review

Provide references using author date format

Bollmann, Saskia. (2022). “Imaging of the Pial Arterial Vasculature of the Human Brain in Vivo Using High-Resolution 7T Time-of-Flight Angiography.” ELife 11 (April). https://doi.org/10.7554/elife.71186.
Chatterjee, Soumick. (2022). “SMILE-UHURA Challenge 2023.” Synapse. December 2, 2022. https://doi.org/10.7303/syn47164761.
Çiçek, Özgün. (2016). “3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation.” Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016, 424–32. https://doi.org/10.1007/978-3-319-46723-8_49.