Influence of deep learning image reconstruction and adaptive statistical iterative reconstruction-V on automated...

The Alberta Stroke Program Early CT Score (ASPECTS) and advances in CT reconstruction play important roles in the neurodiagnostic workflow. This study examines the effect of these reconstruction techniques on automated ASPECTS.

In this retrospective study, 173 patients (median age 79 years, 39% female) with suspected middle cerebral artery infarction underwent non-contrast CT scans reconstructed with Filtered Back Projection (FBP), ASIR-V (30% and 60%), and DLIR (low, medium, and high). Automated ASPECTS were analyzed, with FBP as the reference standard.

Bland-Altman analysis revealed a mean bias with ASIR and DLIR underestimating ASPECTS compared to FBP. This underestimation was less pronounced for ASIR-V 30% (-0.057 ) and DLIR-L (-0.069) than for ASIR-V 60% (-0.126), DLIR-M (-0.121), and DLIR-H (-0.086). The region with the greatest overestimation relative to FBP was M3 (n = 23), while the region with the greatest underestimation was the insular ribbon (n = 51). Regarding the ASPECTS < 6 threshold, most patients were reclassified from ASPECTS ≤ 5 to ASPECTS ≥ 6 with DLIR-M (n = 5), which also showed the strongest agreement with expert consensus (κ = 0.352).

Both ASIR-V and DLIR resulted in only minor underestimation of ASPECTS compared to FBP. However, most patients became eligible for endovascular therapy due to ASPECTS reclassification with DLIR-M. DLIR-M also exhibited the highest agreement with expert consensus for automated ASPECTS. Therefore, careful selection of reconstruction parameters, as well as further optimization and standardization of these techniques, is essential for broader application in stroke imaging.