AI Pinpoints Blocked Brain Arteries in Seconds Using Anatomical Map

When a large vessel occlusion strikes, every minute of delayed treatment translates into lost brain tissue and diminished chances of recovery. These blockages in the brain’s major arteries account for an estimated 24 to 46 percent of acute ischemic strokes and carry a devastating prognosis: fewer than half of affected patients regain functional independence at three months, even with modern endovascular treatment. Now, a team of researchers in Spain has developed an artificial intelligence system that not only detects these critical clots on brain scans but simultaneously identifies exactly which artery is blocked, using the brain’s own plumbing blueprint as a guide. The work, published in the journal Neuroinformatics, promises to give stroke teams faster, more anatomically precise information at the moment it matters most.

The research, led by Valeriia Abramova of the Computer Vision and Robotics Institute at the University of Girona, together with neurologists at Hospital Universitari Dr Josep Trueta, addresses a gap that has limited existing AI stroke tools. Most current detection systems treat the problem as a simple yes-or-no question: is a large vessel occlusion present, or not? But clinicians need more than that. The choice of endovascular strategy differs depending on which vessel segment is occluded. Blockages in the posterior circulation, such as the basilar artery, are more often treated with balloon angioplasty and permanent stents, while occlusions in the anterior circulation typically call for stent thrombectomy or catheter aspiration. Clinical guidelines for posterior circulation occlusions are still evolving, and uncertainty persists over optimal anesthesia and treatment approaches. Knowing the precise vessel involved therefore shapes intervention planning, prognosis estimation, and clinical interpretation in ways a binary alarm cannot.

Commercial software packages such as RapidLVO, Viz-LVO, e-CTA, and StrokeViewer-LVO have brought automated detection into hospitals, but they remain constrained by reduced sensitivity for distal occlusions, vulnerability to imaging artifacts that trigger false positives, and dependence on standardized acquisition protocols that can hamper generalization across institutions. Recent research efforts have pushed toward explicit localization of occlusions, including two-stage convolutional neural network pipelines on four-dimensional CT angiography and self-configuring object detection frameworks applied to maximum intensity projection images. Yet most of these approaches stop at drawing a bounding box around the affected region without linking that region to a specific, named vessel segment. The Girona team’s contribution is to do both tasks at once, within a single unified framework.

Technically, the researchers adapted the state-of-the-art nnDetection framework, a self-configuring three-dimensional object detection method built on the Retina U-Net architecture. This network fuses the RetinaNet one-stage detector with a U-Net encoder-decoder, extracting multi-level image features through convolutional layers with instance normalization and ReLU activation, while lateral and transposed convolutional connections build a feature pyramid. Two sub-networks operate on the pyramid levels: one for classification and one for bounding box regression, trained respectively with binary cross-entropy loss and a generalized intersection-over-union loss, with additional semantic segmentation supervision during training. The classification branch labels each detected occlusion as belonging to one of three clinically crucial segments: the basilar artery, the terminal internal carotid artery, or the M1 segment of the middle cerebral artery. Model training used a fixed patch size of 128 by 128 by 128 voxels, sixty epochs, and stochastic gradient descent with Nesterov momentum, implemented in PyTorch on an NVIDIA GeForce GTX 1080 Ti GPU.

The study’s most distinctive twist lies in how it constrains the search space anatomically. The Circle of Willis is the ring-shaped arterial network at the center of the brain that supplies blood to cerebral tissue, and it encompasses precisely the vessels most often affected by these occlusions. Because manual segmentation of this structure is laborious and ground truth vessel masks were unavailable for the datasets, the team trained a standard nnU-Net segmentation model on data from the TopCoW challenge, a competition dedicated to topology-aware Circle of Willis segmentation in CT and MR angiography spanning diverse anatomical variants. Their segmentation model achieved a Dice score of 0.944 plus or minus 0.026. Applied to the stroke scans, the automatic Circle of Willis segmentation was expanded by 50 voxels in all three dimensions, a margin chosen to guarantee that all occlusions in the dataset fell inside the resulting region of interest.

This anatomical prior enabled a head-to-head comparison of two strategies. In the global approach, the network trained and inferred on the full CT angiography volume. In the local approach, images were cropped to the Circle of Willis-derived region before training and inference. The development dataset comprised 179 CT angiography scans acquired on a Philips Ingenuity scanner at Hospital Dr. Josep Trueta, all containing occlusions annotated with three-dimensional bounding boxes by an expert neurologist. Of these, 143 scans were used for training with five-fold cross-validation, while 36 were held out for internal testing. Class distribution reflected the natural epidemiology of these strokes: the M1 segment dominated at 67 percent of cases, terminal internal carotid artery occlusions comprised 25 percent, and basilar artery occlusions were rare at 8 percent.

The results on the internal test set were strikingly strong for both variants. The global approach achieved a detection sensitivity of 0.92 at 0.08 false positives per image, rising to 0.97 at 0.20 false positives per image, while the local approach matched the same sensitivities at just 0.03 and 0.13 false positives per image respectively. A case-level paired bootstrap analysis at a fixed operating point of 0.1 false positives per image found a mean sensitivity difference of 0.000 with a 95 percent confidence interval spanning negative 0.081 to positive 0.081, confirming no systematic performance gap between the strategies. Each approach missed only a single occlusion. Localization accuracy was equally tight: the mean three-dimensional distance between predicted and ground truth bounding box centers was 1.76 millimeters for both strategies. Interestingly, the full-volume model generated more low-confidence false positives, whereas the region-restricted model produced fewer spurious detections overall.

Classification performance held up nearly as well. On the internal test set, overall accuracy reached 94 percent for the global approach and 91 percent for the local approach, with Cohen’s kappa statistics of 0.88 and 0.82, both indicating almost perfect agreement with expert ground truth. Every basilar artery occlusion was classified correctly by both approaches, a success the authors attribute to the basilar artery’s distinctive, isolated position at the base of the brain, which reduces ambiguity. The main classification shortfall appeared in the terminal internal carotid artery class, where accuracy fell from 89 percent globally to 78 percent locally. Counting detection and classification together, the occlusion was both found and correctly labeled in 33 of 36 cases for the global approach and 32 of 36 for the local one.

Generalization was tested on the independent CODEC-IV benchmark, consisting of 48 CT perfusion-derived CT angiography scans from different scanners and hospitals. Detection sensitivity dropped to 0.71, but much of that decline traced to a mismatch in annotation conventions: the benchmark’s ground truth boxes were uniformly small and pinpointed the occlusion site, while the Girona training boxes captured the entire clot extent. When the true positive threshold was relaxed to 2 percent intersection over union, sensitivity climbed to 0.90 for the global approach and 0.88 for the local one, with mean center-to-center distances of roughly 2.4 millimeters confirming that the models were accurately placing their predictions even when box dimensions diverged. On a supplementary inference-only evaluation of the IACTA-EST challenge dataset, which included 51 occlusion-negative cases, the global model achieved an area under the curve of 0.95 in separating positive from negative cases, against 0.84 for the local model, hinting that a well-chosen confidence threshold could suppress false alarms in clinical triage.

Perhaps the most consequential number is temporal. Restricting inference to the Circle of Willis region made the local approach approximately 3.3 times faster than the global one, cutting mean processing time per case from about 227 seconds to roughly 89 seconds. In acute stroke care, where treatment delays of minutes measurably worsen neurological outcomes, that speedup matters even before accounting for the reduced computational cost of running on modest hospital hardware. The authors acknowledge limitations, including a single-center training cohort from one scanner, a small number of basilar artery cases, and class imbalance mirroring natural disease prevalence. Nonetheless, by coupling a deep object detection framework to an automatically derived vascular landmark, the study demonstrates that anatomical knowledge can be baked directly into machine learning pipelines, delivering detection and vessel-level diagnosis in one pass and pointing toward AI assistants that fit realistically into the breakneck rhythm of a stroke unit.

Subject of Research: Deep learning detection and vessel-level classification of large vessel occlusions in brain CT angiography guided by Circle of Willis localization

Article Title: Circle of Willis-Guided Localization for Simultaneous Detection and Classification of Large Vessel Occlusions in Brain CTA

Article References: Abramova, V., Oliver, A., Lal-Trehan Estrada, U. M., Hamadache, R. E., Martínez Arias, P., Freixenet, J., Terceño, M., Silva, Y., & Lladó, X. (2026). Circle of Willis-Guided Localization for Simultaneous Detection and Classification of Large Vessel Occlusions in Brain CTA. Neuroinformatics, 24(4), Article 62. https://doi.org/10.1007/s12021-026-09817-x

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09817-x

Keywords: large vessel occlusion, stroke, CT angiography, Circle of Willis, deep learning, nnDetection, neuroinformatics, medical imaging, endovascular treatment, brain imaging, artificial intelligence, stroke diagnosis