Logo image
Acute Infarct Core Volume Estimation on Noncontrast Computed Tomography With a Deep Learning Algorithm
Journal article   Open access   Peer reviewed

Acute Infarct Core Volume Estimation on Noncontrast Computed Tomography With a Deep Learning Algorithm

Santiago Ortega-Gutierrez, Juan Vivanco-Suarez, Aaron Rodriguez-Calienes, Victor Salvia, Milagros Galecio-Castillo, Mahmoud Dibas, Yujing Lu, Alba Garcia Rey, Cristian Marti, Leonardo Tanzi, …
Stroke: vascular and interventional neurology, Vol.5(2), e001509
03/2025
DOI: 10.1161/SVIN.124.001509
PMCID: PMC12671631
PMID: 41573179
url
https://doi.org/10.1161/SVIN.124.001509View
Published (Version of record) Open Access

Abstract

BACKGROUND A simplified patient selection paradigm with noncontrast computed tomography (NCCT) can reduce the time to reperfusion and widen the eligibility of acute ischemic stroke large vessel occlusions (LVOs) for endovascular therapy. The objectives of this article are (1) to develop, train, and internally validate a deep learning algorithm that estimates baseline infarct core volume (ICV) on NCCT in anterior circulation LVO patients, and (2) by using an external set, to ascertain how this algorithm's (aICV‐NCCT) predictive performance compares with Alberta Stroke Program Early Computed Tomography Score‐NCCT and ICV‐CT perfusion in its capacity to estimate the final infarct volume established on diffusion‐weighted magnetic resonance imaging at 24‐ to 48‐hour follow‐up. METHODS In the first phase, stroke activations with baseline NCCT and CT angiography were used to train an aICV‐NCCT. The algorithm was then internally validated using intraclass correlations and Intersection over Union. In the external set, patients with LVO treated with endovascular therapy achieving modified Thrombolysis in Cerebral Infarction score ≥2b and available baseline NCCT, CT angiography, and CT perfusion were included. RESULTS A total of 2858 studies of patients with stroke alerts were used for training (80%) and internal validation (20%). We obtained a high correlation (intraclass correlation coefficient, 0.78; CI, 0.73–0.83) and an acceptable Intersection over Union of 0.24 on the internal validation set. The external set consisted on 230 patients with an LVO. When predicting final infarct volume on the external set, our aICV‐NCCT was similar to ICV‐CT perfusion (intraclass correlation coefficient, 0.50 versus 0.54; P = 0.764) and Alberta Stroke Program Early Computed Tomography Score‐NCCT (rs, −0.41; P = 0.436). CONCLUSION In this study, we developed and validated a deep learning algorithm that demonstrates an at least equivalent performance to CT perfusion in estimating core volume on acute stroke imaging studies in patients with suspected anterior circulation LVO strokes. The algorithm's robust performance holds significant potential in settings with limited access to advanced imaging technologies across diverse healthcare environments.
Computed Tomography Machine Learning Software Stroke infarct radiographic imaging UIOWA OA Agreement

Details

Metrics

Logo image