Prof. Peter Chang, MD

When a patient comes in for suspected stroke, a number of key decisions are made based on imaging. If a non-contrast head CT excludes acute hemorrhage, the patient is triaged and evaluated to identify an occlusion or thrombus in the proximal Circle of Willis. Patients with the right type of potentially treatable thrombus are then evaluated and characterized for the extent of ischemic core via ASPECTS score or CT perfusion. Each of these steps represents a potential opportunity for AI-enabled quantitative evaluation.

Here are two applications in use at our hospital. The one on the left evaluates acute hemorrhage, while the one on the right evaluates large vessel occlusion. An AI algorithm goes through these busy work lists and identifies the positive patients in a matter of seconds. If you click on any of the positive patients and scroll through the exam, you can confirm the presence of hemorrhage or large vessel occlusion.

Such findings are not difficult for a trained expert to identify, but these triage applications do it much faster. And the AI algorithm can send notifications in real time of positive findings – to your mobile phone or hospital text page system – further enhancing the ability to alert and attend to patients in need of evaluation.

In the hemorrhage application, one algorithm identifies suspicious regions in the brain. A second algorithm carefully evaluates these regions for the presence or absence of true hemorrhage. Then a third algorithm evaluates the true hemorrhage to create very precise segmentation masks for hemorrhage volume.

An initial series of 10,000 exams not used for algorithm training yielded a sensitivity of over 95% with just 26 missed cases out of 900 positive exams. Moreover, the algorithm was within about two percent of the average true hemor- rhage volume, while a more conventional ABC/2 measurement dependent on manual 2D measurement overestimated the true hemorrhage volume by 20%.

Our hemorrhage application algorithm can pick out tiny hemorrhages that are sometimes missed in busy clinical practices by expert humans. However, it isn’t perfect. It sometimes misses small-volume subarachnoid hemorrhages and produces false negatives and positives due to things like artifacts mimicking high density and general forms of pathologic high density that simply are not hemorrhages.

This was one of the first AI tools to qualify for CMS reimbursement as part of the New Technology Add-on Payment announced in late 2020 – because it demonstrated improvement in patient outcomes over the baseline. When used in large vessel occlusion detection for any inpatient who accrues charges beyond what is typically reimbursed for standard care, the hospital is eligible for up to an additional 1,040 dollars of payment.

The other method for assessing core infarct is CT perfusion. We have some early AI-based prototypes for this task, but well-established non-Deep Learning methods are perfectly satisfactory.

In summary, each triage-based imaging decision that one needs to make in a stroke patient – whether detection of hemorrhage, detection of proximal large vessel occlusion, or characterization of ischemic core – is an opportunity for Deep Learning AI to help expedite and improve the overall clinical workflow.

See more topics in this webinar.