When a major blood vessel in the brain is blocked, every delay can matter. This study asks whether a computer can spot that blockage on a routine scan—and also judge how well blood is finding another route around it. Stroke affects millions of people every year, causing deaths and lasting disabilities.

A stroke caused by a blocked major brain vessel is among the most severe forms, and about one in three ischemic strokes is caused by this kind of blockage. For acute ischemic stroke with a large-vessel blockage, the standard treatment is mechanical thrombectomy, a procedure used to remove the blockage and restore blood flow.

The scan used here is a CTA scan. The computer was asked to find LVO and judge the collateral blood flow around it. It had to do both jobs from a single scan, then face scans from a separate source to test whether it could work beyond the data used to build it.

The computer used a three-dimensional learning system that finds objects in medical images. Its structure chooses an arrangement suited to the data instead of relying on one fixed design. The model was built on the Retina U-net and used a similar network topology as its underlying structure in this study.

The model combines detailed and broad views of each brain scan, then uses them to locate the middle cerebral artery and decide whether a major blockage is present. When a blockage is found, an additional pathway identifies the affected side and judges the strength of alternative blood flow.

On scans from a separate center, the model correctly classified 339 of 345 scans, giving an accuracy of 98.26 percent for finding the blockage. Its ability to identify affected and unaffected scans was also strong, with sensitivity and specificity both reported above 96 percent.

The model was trained using scans from different centers, scanners, scan procedures, and contrast phases. On independent data, it still achieved accuracy above 98 percent for identifying the blockage. It also agreed strongly with the specialists when assigning backup-blood-flow scores, matching or exceeding the reliability of individual radiologists compared with their shared judgment.

There is an important caution: the separate test set was relatively small, even though the full study sample was considerably larger. The backup-flow scores came from human radiologists, so their own judgments and biases were passed on to the computer. In the end, the system could identify a major vessel blockage on a single scan and mark its location with a clear box.

It could also provide a score for backup blood flow, bringing two urgent stroke checks into one automated reading. The computer identified these dangerous blockages with more than ninety-eight percent accuracy on scans from a separate center, while also closely matching specialists when judging backup blood flow.

That could help urgent stroke decisions happen more reliably.