‘Minutes matter’: Deep learning quickly spots stroke-causing blockages on CTA scans

A new deep learning model built from scratch can help radiologists quickly pinpoint arterial blockages responsible for stroke, likely speeding up the time to begin vital treatment.

Brown University radiologists detailed their development Tuesday in Radiology, noting such blockages, known as large vessel occlusions (LVOs), account for an overwhelming percentage of ischemic strokes. The team used a large sample of CT angiography (CTA) exams, along with a newer multiphase CTA approach to enhance detection and improve diagnoses for these patients.

And when dealing with this most common type of stroke, time is of the essence, explained lead author Matthew T. Stib, MD, a radiology resident at the Providence, Rhode Island, medical school

“Minutes matter in this time-sensitive diagnosis,” Stib added in a Sept. 29 statement. “Every minute that we reduce the time to recanalization extends the patient’s disability-free life by a week.”

Currently, CTA exams are considered the gold standard for detecting large vessel occlusions. And although radiologists are highly capable of spotting them, they aren’t always readily available.

To help avoid any detrimental delays in care, the researchers created their algorithm and trained it to spot and distinguish LVOs from other conditions. Stib et al. used preprocessed CTA exams and multiphase CTAs. The latter offers more comprehensive information.

After putting the platform to the test on 62 multiphase exams, it notched a 100% specificity, detecting all 31 LVOs. The authors said the results were “quite promising,” pointing out the improvement from the 77% specificity rate achieved using traditional CTA scans.

Next up the researchers plan to test their algorithm in real-world scenarios to determine if it's ready to assist medical centers and hospitals.

“This algorithm is not replacing the ability of radiologists to do their job; rather, it’s trying to speed up the time to diagnosis,” Stib explained. “So if the radiologist isn’t around or there is a large workflow that is preventing someone from looking at the exam results quickly, there will be an alert that says an occlusion may be present and someone should look at this. That’s where the value is in this kind of a model.”