Muhammad Taimoor Khan, MD
As a vascular neurology fellow, an understanding of the automated tools available for immediate diagnosis of large vessel occlusion (LVO), estimation of core and penumbra in the context of treatment decision making has become critical in the era of endovascular therapy. In the article “Machine Learning–Enabled Automated Determination of Acute Ischemic Core From Computed Tomography Angiography,” the authors developed, validated, and reported a deep learning method called “DeepSymNet” that evaluates for ischemic core volume using computed tomography angiogram (CTA) source images. The study included patients with acute ischemic stroke and stroke mimics with CTA and CT perfusion (CTP) using the RAPID software and trained their algorithm against RAPID CTP determinations of ischemic core.
From 297 included patients, 224 (75%) had acute ischemic stroke, of which 179 (60%) had large vessel occlusion. The mean RAPID CTP based ischemic core volume was 23±42 cc. DeepSymNet learned to identify vessels on CTA, detect LVO autonomously and ischemic core of less than or equal to 30 cc and 50 cc with AUC 0.88 and 0.90 (ischemic core ≤30 mL and ≤50 mL) to CTP-RAPID ischemic core volume both in early, 0 to 6 hours and late 6-24 hours time windows. (AUCs 0.90 and 0.91, ischemic core ≤50 mL).
There are other products available, both commercial and non-commercial, with the ability of machine learning to automate analysis needed to screen for endovascular therapy. The uniqueness of DeepSymNet is predictions of ischemic core from CTA, as that can be of great utility in screening stroke patients in hospitals before transferring them to thrombectomy-capable centers.
The authors explain the limitations given their single-center experience, lack of algorithm training against more precise core modality like MRI, and absence of penumbra measurements at this point. Given the initial phase of this product, we believe other limitations are how scalable it will be to other centers, different scanners and imaging protocols in particular.
I would like to thank Dr. Marie Luby, PhD, for guiding me with selecting and preparing this blog post.
