What the fuck good does estimating core infarct do? Regardless of core infarct size you are still required to get to 100% recovery! GET THERE!
Article Commentary: “Automatic Ischemic Core Estimation Based on Noncontrast-Enhanced Computed Tomography”
In the high-stakes world of acute stroke care, the clock ticks relentlessly. Stroke code decision-making is a balance between the urgency of intervention and the value of comprehensive diagnostic information. Knowing the extent of ischemic core infarct, tissue that is irretrievably lost, provides better understanding of the risks and benefits when offering thrombolytics or mechanical thrombectomy. Magnetic resonance imaging (MRI) may be the gold standard for core estimation, but computed tomography perfusion (CTP) is more commonly used as it is faster and more widely available.1 Attempts to skip additional imaging beyond noncontrast-enhanced computed tomography (NCCT) by using the Alberta Stroke Program Early CT Score (ASPECTS) have been limited by its imprecise 10-point scale and notoriously poor interrater reliability. In this article, Nishi et al lay a foundation to make ASPECTS obsolete by using a fully automated machine learning–based technique for NCCT-derived ischemic core.
This multicenter retrospective study trained a machine learning model on pretreatment images from 272 anterior circulation ischemic stroke patients who underwent both NCCT and MRI. The NCCT images were manually labeled to include early ischemic core segmentation by 2 neurointerventionalists using the corresponding MRI diffusion-weighted images (DWI) as a guide. The final model is fully automatic, requires no preprocessing, and can be processed within 2.5 seconds. This study represents a new state-of-the-art in ischemic core segmentation using only NCCT. Compared to the MRI-based reference, the model achieved strong correlation (Pearson r=0.91, P<0.01) with a median core infarct volume difference of 4.7 mL (interquartile range, 0.8–12.4 mL). The dataset was limited to mostly small ischemic core volumes, which resulted in a few notable underestimations. There was reliable performance both in the early time window (0–4.5 hours, r=0.91, P<0.01) and the late time window (4.5–24 hours, r=0.91, P<0.01) from symptom onset to imaging. Previous attempts have reported Pearson correlation coefficients ranging from 0.44 to 0.76. The performance of the model was externally validated across different scanning protocols, vendors, and tube currents, with consistent performance.
Despite its promising results, the real-world utility of DWI-volume prediction alone is unknown. MRI additionally provides the fluid-attenuated inversion recovery (FLAIR) sequence, and it is the mismatch between abnormal hyperintensity on DWI and the lack thereof on FLAIR that has been used to guide successful reperfusion therapy.2 CTP provides volumes for both core and penumbra, tissue at risk, which allows a calculation of the maximum amount of benefit that intervention could offer. While automatic NCCT-derived ischemic core estimation may not replace CTP or MRI, it could still provide diagnostic or prognostic value without having to wait for those additional studies. Future studies should be designed to precisely define that value, potentially using an updated model trained on a larger, more diverse dataset.
