Stroke atlas of the brain: A voxel-wise density-based clustering of infarct lesion topographic distribution
Yanlu Wang1,2, Hadrien Van Loo3, Julia Juliano4, Sook-Lei Liew5,6, Alexander McKinney IV7, and Sam Payabvash8

1Clinical Sciences, Intervention and Technology, Karolinska Institute, Sollentuna, Sweden, 2Medical Radiation Physics and Nuclear Medicine, Karolinska University Hospital, Stockholm, Sweden, 3Medical Radiation Physics and Nuclear Medicine, Karolinska University Hospital, Huddinge, Sweden, 4University of Southern California, Los Angeles, CA, United States, 5Viterbi School, Department of Biomedical Engineering, Chan Division of Occupational Science and Occupational Therapy, University of Southern California, Kek School of Medicine, Los Angeles, CA, United States, 6Department of Neurology USC Stevens Neuroimaging and Informatics Institute, Division of Biokinesiology and Physical Therapy, University of Southern California, Keck School of Medicin, Los Angeles, CA, United States, 7Department of Radiology, University of Minnesota, Minneapolis, MN, United States, 8Yale Medicine, New Haven, CT, United States


In stroke patients, both infarct volume and location affect functional outcome; however, infarct topography is far less commonly incorporated in prognostic models, given the complexity of assessing infarct topographic distribution. In this study, we applied data-driven density clustering analysis, using the OPTICS algorithm, on 793 infarct lesions from 438 stroke patients to devise a “stroke-atlas of the brain” stratifying brain voxels likely to infarct together. This atlas can help with differentiation of infarct lesions in clinical practice, assess topographic distribution of infarct in prognostic models for stroke patients, or be applied for defining regional infarct thresholds in CT/MR perfusion maps.


In this study, we applied data-driven density clustering analysis to stroke legion data, to generate a probability-varying stroke atlas of the brain – depicting voxels/regions that are likely to infarct simultaneously. The atlas topology can change depending on the probability thresholds set, allowing visualization of both small regions of extremely high infarct probability regions, to very large, but less probabilistically stringent simultaneous infarct regions.

Materials and Methods

Two different datasets were used in current study: (1) patients with acute stroke and DWI scan performed within 24 hours of symptom onset; and (2) patients with chronic infarcts. In the cohort (n=238) with acute stroke, the infarct lesions (n=405) were manually segmented on DWI scans; whereas, in the cohort with chronic infarcts, the lesions (n=393) were manually segmented on 3D T1-weighted images. The data were driven from 11 centers worldwide with 17 different scanners with the approval of the institutional review boards at respective institutes. All manual segmentations were performed and/or supervised by neuroradiologists using MRIcro software. All infarct lesions were saved as binary masks and coregistered to the standard MNI template using a 12-parameters affine transformation. The l2-norm of the voxel-wise joint probability was used as distance metric for the clustering algorithm. Density Clustering was applied using the full OPTICS algorithm, to preserve the full reachability plot. This allows the extraction of sets of clusters for different reachability thresholds extremely fast, without performing the full clustering algorithm, which is very computation and memory intensive, each time.


The resulting stroke map can be viewed at different reachability, or “probability” thresholds. Clusters formed at a given threshold depict infarct voxels/regions that are equally likely between the clusters. Figure 1 shows the set of 20 clusters first formed when gradually increasing the intra-cluster homogeneity threshold. This depicts a broad stroke map covering most of the brain and conforms to the major arterial supply territories. When the intra-cluster homogeneity constraint is gradually tightened, the number of resulting clusters increases while the average size of the clusters decreases (Figure 2). Figure 3 depicts the set of clusters (n=206), when the intra-cluster homogeneity constraint is tightened. Given a voxel/region within a cluster in this map has succumbed to infarct, the probability that another voxel/region within the same cluster will also succumb to infarct is much higher compared to Figure 1. At these reachability thresholds, there are simply not enough visually distinct color codes to feasibly visualize all the clusters at once. To feasibly visualize clusters at low reachability thresholds, one may only view parts of the reachability plot, corresponding to a subdivision of larger clusters at higher reachability thresholds. Figure 4 depicts one such subdivision, where the largest cluster from Figure 1 is subdivided into approximately 20 clusters of >1ml size (corresponding to 37 voxels), omitting smaller clusters.


Visualization of the results of density clustering based on different reachability thresholds is challenging. We are currently devising schemes to visualize the entire clustering structure with varying intra-cluster likelihood constraints in a concise and intuitive fashion. One may speculate that infarct clusters in the stroke atlas may represent arterial perfusion territories. Thus, in case of an atherosclerotic or embolic arterial occlusion, the brain regions/voxels supplied by specific arterial branch(es) tend to infarct together. Nevertheless, the regional boundaries and topographic distribution should be interpreted with caution given the amount of anatomical variation in arterial supply pattern of the brain.

There are many potential applications for the proposed stroke atlas. Topographic delineation of brain regions that are likely to infarct together can help radiologists in clinical practice with differentiation of ischemic infarct from say hypercellular metastasis – which can have restricted diffusion – in patients with cancer. The proposed brain parcellation method can also be applied to evaluate topographic distribution of infarct for multivariate prognostic models in stroke patients. Finally, reverse coregistration of the stroke atlas onto brain CT/MR perfusion maps can be used for calculation of regional infarct core/penumbra thresholds.


We have successfully applied density clustering to 793 lesions to delineate brain regions (voxels) likely to infarct simultaneously in stroke patients. Varying the reachability threshold allows us to tune the output from small, but probabilistically homogenous regions, to large brain areas, but less homogenous in simultaneous infarct likelihood. The proposed brain parcellation map may represent meticulous arterial perfusion territories that tend to infarct simultaneously in context of thromboembolic event. Such a stroke-atlas of the brain can help with clinical differentiation of infarct lesions, assess infarct topology in multivariate prognostic models, or refine CT/MR perfusion maps based on regional thresholds.


No acknowledgement found.


1. Payabvash S, Taleb S, Benson JC, McKinney AM. Interhemispheric Asymmetry in Distribution of Infarct Lesions among Acute Ischemic Stroke Patients Presenting to Hospital. J Stroke Cerebrovasc Dis. 2016;25(10):2464-9.

2. Liew SL, Anglin JM, Banks NW, et al. A large, open source dataset of stroke anatomical brain images and manual lesion segmentations. Sci Data. 2018;5:180011.

3. Payabvash S, Benson JC, Tyan AE, Taleb S, McKinney AM. Multivariate Prognostic Model of Acute Stroke Combining Admission Infarct Location and Symptom Severity: A Proof-of-Concept Study. J Stroke Cerebrovasc Dis. 2018;27(4):936-44.

4. Payabvash S, Taleb S, Benson JC, McKinney AM. Acute Ischemic Stroke Infarct Topology: Association with Lesion Volume and Severity of Symptoms at Admission and Discharge. AJNR Am J Neuroradiol. 2017;38(1):58-63.

5. Payabvash S, Souza LC, Wang Y, et al. Regional ischemic vulnerability of the brain to hypoperfusion: the need for location specific computed tomography perfusion thresholds in acute stroke patients. Stroke. 2011;42(5):1255-60.


Formation of the 20 largest clusters. These large clusters are the least homogenous in intra-cluster reachability, but in return, covers most of the brain. This map can be considered a broad stroke map, which conforms to the major arterial supply territories. Color coding for clusters is arbitrary.

Cluster statistics with varying reachability threshold. (A) The number of clusters formed decreases with increased reachability threshold. (B) The mean cluster size of formed clusters increases with reachability threshold. This is an understandable consequence of (A). The increase in mean cluster size is only slight at low reachability thresholds and escalates at higher thresholds. (C) The percentage of voxels in the brain that is included in a cluster. As reachability threshold increases, the homogeneity condition is relaxed, and thus more voxels can partake in a cluster. The steady increase in voxel percentage flattens slightly prior to reaching 100% participation.

Formation of approximately 200 clusters (n=206), when the intra-cluster homogeneity constraint is tightened. This depicts is more finely detailed stroke map. Compared to figure 1, the probability that another voxel/region within the same cluster will also succumb to infarct is much higher. Visualization problems occur at these cluster numbers however, since there are not enough visually distinct color codes to clearly display this map. Color coding for clusters is arbitrary.

Subdivision of the largest cluster (nr. 1) in figure 1 into its 20 largest sub-components, where we can retain clarity in visualization through visually distinct colors. This corresponds to only displaying parts of the full reachability plot. Color coding for clusters is arbitrary. (A) Full reachability plot. Colored columns indicate clusters formed (see fig. 1). (B) Reachability plot of the subdivision of the leftmost cluster in (A, left-most cluster) (C) The results of the Sub-division, excluding sub-clusters which are less than 1ml in volume. Visually distinct color coding of all sub-divided clusters is now possible.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)