Purpose: Computer-aided diagnosis (CAD) appears promising in early ischemic change detection computed tomography (CT). This study aimed to compare the performance of two new CAD systems (Frontier ASPECTS Prototype and Brainomix) with two experienced readers in selected patients with suspected acute ischemic stroke.
Methods: Retrospectively, non-contrast brain CTs of 150 patients suspected for acute middle cerebral artery ischemia were analyzed with respect to ASPECTS first separately, than in consensus by two senior radiologists, and by use of Frontier and Brainomix. Besides the fully automatic Frontier and Brainomix readings (Frontier_1, Brainomix_1), readings adjusted for the affected brain side (known by CT angiography or clinical presentation, Frontier_2, Brainomix_2) were assessed. Statistical analysis was performed by intraclass correlation and Bland-Altman statistics.
Results: The score-based ASPECTS readings of Brainomix_1, Brainomix_2, both radiologists, and the expert consensus reading correlated highly (r = 0.714 to 0.841; always p < 0.001), whereas Frontier_1 and Frontier_2 correlated only lowly or moderately with both radiologists, the expert consensus reading, and Brainomix (r = 0.471 to 0.680; always p < 0.001). Bland-Altman analysis revealed lower mean ASPECT difference and standard deviation of difference for Brainomix_2 (mean difference = -0.2; SD = 1.15) compared to Frontier_2 (mean difference = 1.2; SD = 1.76). Correlation of region-based ASPECTS reading with the expert consensus reading was moderate for Brainomix_2 (r = 0.534), but only low for Frontier_2 (r = 0283; always p < 0.001).
Conclusion: We found high agreement in ASPECTS rating between both radiologists, expert consensus reading, and Brainomix, but only low to moderate agreement to Frontier.
Keywords: ASPECTS; Artificial intelligence; Brain computed tomography; Computer-aided diagnosis; Early ischemic change detection.