Topology-Aware Anatomical Segmentation of the Circle of Willis for CTA and MRA

The Circle of Willis (CoW) is an important anastomotic network of arteries connecting the anterior and posterior circulations of the brain, as well as the left and right cerebral hemispheres [1]. The CoW is located at the skull base and is the joining area of major arteries to the brain. Due to its centrality, the CoW is commonly involved in pathologies like aneurysms and stroke. Clinically, the vascular architecture of the CoW is believed to impact the occurrence and severity of stroke [2, 3]. An accurate characterization of the CoW is therefore of great clinical relevance.

However, clinicians have articulated an unmet demand for efficient software tools to analyze and compare the angio-architecture of the CoW. Today, assessing the anatomy and vascular components of the CoW from medical angiography images is still an expert task and time-consuming. Furthermore, the CoW naturally has many variants of which certain principal artery components are hypoplastic or absent. It is estimated that only around half of our population has a complete CoW [4, 5]. These anatomical CoW variants should not be misinterpreted as a vascular disease, and it is not an exception to see the CoW vary markedly from person to person. An automated and personalized CoW vascular characterization will be of significant interest to both the clinical and the research communities.

Here, we propose the first public challenge on CoW angio-architecture extraction and brain vessel segment annotation on two common angiographic imaging modalities, namely magnetic resonance angiography (MRA) and computed tomography angiography (CTA). Although there have been two publicly available datasets on MRA
modality, the CASILab and IXI datasets [6,7], their MRA scans are acquired from fairly old machines (from the year 2004 to 2006). More importantly, very limited annotations were provided if any. Annotated dataset on the other important modality, CTA, is even rarer. Thus, we identify a gap in good quality anatomical annotations, newer
imaging datasets, and datasets of dual modalities. By releasing a public dataset covering both CTA and MRA with careful annotations on the vasculature, our challenge can foster research and better benchmark comparisons on the CoW vascular segmentation results.

The aim of the challenge is threefold; first, to extract the CoW angioarchitecture from 3D angiographic imaging; second, to automatically annotate the vessel components; and third, to characterize the CoW topologically. We release a new dataset of joint-modalities, CTA and MRA of the same patient cohort, both with annotations of the underlying anatomy of CoW. There is a temporal relationship between the two modalities from the same patients in the form of follow-up scans. The inclusion criteria of the patient cohort is that at least one of the modalities allow for a diagnosis of the underlying CoW anatomical and geometric characterization. Joint-modalities from the same patient will serve as additional reference and provide supplementary anatomical information on the CoW vessels. We believe our temporal joint-modality annotations on MRA and CTA can ensure good quality anatomical annotations. Such a joint-modality dataset also opens doors to other medical imaging research questions such as registration and unpaired modality network design. Our challenge has two tracks for the same segmentation task with one track for each modality. We leverage both modalities during our annotation. And participants can leverage whichever modality they want, both CTA and MRA, and choose to tackle the task for either modality.

A technical emphasis of this challenge is on topology-aware segmentation. The extracted vessels should retain the topology of the underlying anatomy. We especially hope to raise awareness on the importance of evaluating performance beyond pixel-based or volumetric metrics. In particular, we will evaluate the segmentation performance on topology-based, junction-based, and graph-based metrics. The objects of interest will include centerlines, bifurcation points, and geometric shapes. Segmentation gives a  complete description of the geometry of the vasculature. Extracting features such as centerline, radii, bifurcation points from segmentation has been frequently done. The challenge aims for vascular characterizations that capture the underlying topology and geometric variability of the CoW.

We believe this is a timely and new challenge that can be of benefit to both clinicians and medical imaging researchers. An automated CoW characterization on vessel anatomical labelings and topological properties can lead to many interesting downstream tasks and applications. Apart from enabling neurovascular disease management and personalized treatment planning, this challenge can also impact research and discovery. There are gems encoded in the angio-architecture of CoW such as hemodynamic implications and flow analysis. Profiling CoW variants can also reveal demographic geometric risk factors for vascular pathologies. An accurate CoW anatomical segmentation helps in development of more complex models, more efficient quantification, and reduction of cognitive workload and more consistent labeling procedure. Based on the  clinical impact, data value, and research potential, we look forward to organizing this challenge (and its repeats) and welcoming submissions.

References:
[1] AG, Osborn. "Osborn’s brain: imaging, pathology, and anatomy." Salt Lake City, Utah: Amirsys (2013): 932-940.
[2] Chuang, Yu-Ming, et al. "Configuration of the circle of Willis is associated with less symptomatic intracerebral hemorrhage in ischemic stroke patients treated with intravenous thrombolysis." Journal of critical care 28.2 (2013): 166-172.
[3] van Seeters, Tom, et al. "Completeness of the circle of Willis and risk of ischemic stroke in patients without cerebrovascular disease." Neuroradiology 57.12 (2015): 1247-1251.
[4] Krabbe-Hartkamp, Monique J., et al. "Circle of Willis: morphologic variation on three-dimensional time-of-flight MR angiograms." Radiology 207.1 (1998): 103-111.
[5] Iqbal, S. "A comprehensive study of the anatomical variations of the circle of willis in adult human brains." Journal of clinical and diagnostic research: JCDR 7.11 (2013): 2423.
[6] Bullitt, Elizabeth, et al. "Vessel tortuosity and brain tumor malignancy: a blinded study1." Academic radiology 12.10 (2005): 1232-1240.
[7] IXI - Information eXtraction from Images. https://brain-development.org/ixi-dataset/