Author: Matthew S Brown, John M. Hoffman, William Hsu, Grace Hyun Kim, Michael F. McNitt-Gray, Spencer Harrison Welland, Anil Yadav 👨🔬
Affiliation: Department of Bioengineering, UCLA, David Geffen School of Medicine at UCLA, UCLA Department of Radiology 🌍
Purpose: Non-contrast CT (NCCT) is frequently used in initial evaluation of suspected stroke to rule out intracerebral hemorrhage. Quantitative scoring systems like the Alberta Stroke Program Early CT Score (ASPECTS) have been developed to quantify the severity of stroke. ASPECTS, as well as the volume of hypodense brain tissue is used for determining eligibility of thrombolytic treatment. Artificial Intelligence (AI)-based tools have been developed to automate measurement of ASPECTS and hypotenuse volume. The purpose of this work was to evaluate the effect of CT reconstruction kernel and slice thickness on AI-based hypodense volume and ASPECTS measurements.
Methods: The NCCT series image data of 25 patients scanned with a CT stroke protocol at our institution were reconstructed using a smooth, medium, medium sharp, and sharp kernel (Siemens H10s, H40s, H60s, and H70h) and 1.5mm, 3.0mm and 5mm slice thickness to create a reference condition (H40s/5mm) and 11 non-reference reconstruction conditions. Each reconstruction was analyzed using the Brainomix e-Stroke software (Brainomix, Oxford, England) which yielded an ASPECTS value. The Kruskal-Wallis (KW) test was used to for statistical hypothesis testing, followed by Dunn's post hoc test.
Results: In the reference condition (H40s/5.0mm), mean hypodense volume was 12.0±6.5 ml (mean±SD) and mean ASPECTS was 9±1.8. The e-Stroke software failed to return ASPECTS values for all cases under the H70h/1.5mm condition. The KW test revealed a significant difference between reference and non-reference hypodense volume (p < 0.001) but not ASPECTS (p = 0.9). Dunn's test of hypodense volume measurements revealed significant differences in images reconstructed with the H60s and H70h kernels, and 1.5mm slice thickness.
Conclusion: AI-based measurement of hypodense volume in NCCT can be significantly affected by CT reconstruction kernel and slice thickness. Future work will explore methods to improve reproducibility of hypodense volume across CT reconstruction conditions.