Abstract Body:
Introduction: Automated imaging analysis tools are increasingly used in clinical decision-making for stroke. Rapid ASPECTS (iSchemaView, Menlo Park, CA) assists physicians by automatically calculating Alberta Stroke Program Early CT Scores (ASPECTS) and reducing inter-reader variability. To understand why the tool’s performance in real-world settings sometimes varies compared to published literature, we investigated how different imaging acquisition protocols affect its performance.

Materials & Methods: Consecutive code stroke NCCT scans with thin (1.25 mm slice; 0.625 mm spacing) and thick (5.0 mm slice; 3 mm spacing) series were collected from a retrospective database between February 2020 and May 2021. Ground truth ASPECTS reads were collected from radiology reports, which neuroradiologists determined in real-time. Automated reads were obtained using Rapid ASPECTS 1.0 and 3.0 (iSchemaView, Menlo Park, CA). Agreement between automated and manual reads was defined as ASPECTS scores within two points.

Results: A total of 682 cases were included in this analysis. 67 cases were excluded for technical inadequacy (hemorrhages, tumors, and artifacts). A review of the source imaging revealed that many cases had thick overlapping slices and incorrect head positioning (neck extended instead of the standard neutral position). These cases required significant tilt correction to align the patient data with the Rapid ASPECTS regions template. These corrections led to partial voluming artifacts, which caused lower Hounsfield unit (HU) values and ASPECTS scores. When adjusting protocols from thick to thin slices, agreement between ASPECTS V1 and manual reads improved from 85% (581/682) to 89% (606/682). ASPECTS V3 showed further improvement, with agreements of 91% (619/682) and 95% (648/682) for thick and thin slice scans, respectively.

Conclusion: The combination of neck extension head positioning and thick overlapping slices caused partial voluming artifacts, resulting in artificially low ASPECTS scores on automated software. Our findings indicate that adjusting imaging protocols and working with the AI provider can enhance an algorithm's accuracy. To ensure that commercially available automated analysis tools deliver accurate results, it is crucial to follow the recommended imaging acquisition protocols.