Abstract
BACKGROUND: The choice of arterial input function (AIF) during post-processing of computed tomography perfusion (CTP) imaging can strongly influence perfusion maps. At present, there is no consensus on the optimal site for AIF selection. Since CTP often overestimates hypoperfused tissue, some patients with acute ischemic stroke (AIS) may face misdiagnosis and unnecessary treatment. This study aimed to improve the accuracy of hypoperfusion assessments through artificial intelligence to automatically modify the selection of AIFs. METHODS: We retrospectively analyzed 35 patients with AIS caused by unilateral anterior circulation obstruction who did not undergo thrombolysis or thrombectomy. Each patient underwent an emergency "one-stop" computed tomography (CT) scan, including non-contrast CT, CT angiography, and CTP, followed by magnetic resonance imaging (MRI) within 10 days. AIF was measured at two sites: (I) a normal large artery (AIF(NLA)); and (II) a collateral arteriole of the middle cerebral artery (MCA) adjacent to the ischemic lesion (AIF(AIL)). Hypoperfusion volumes were compared with final infarct volumes (FIVs) defined on magnetic resonance (MR) diffusion-weighted imaging (DWI). Agreement was assessed using Bland-Altman analysis, Spearman correlation, Dice similarity coefficient, Hausdorff distance (HD), positive predictive value (PPV), true negative rate (TNR), false negative rate (FNR), and overall accuracy. RESULTS: A total of 35 eligible patients were analyzed. The mean absolute error (MAE) for AIF(NLA) was 55.66 mL, compared with 20.26, 34.69, and 44.06 mL when measured from AIF(AIL) with 4-, 6-, and 8-second delays, respectively. Hypoperfusion volumes based on the AIF(AIL) with a 4-second delay did not differ significantly from FIVs (P=0.43), whereas other methods showed significant differences (all P<0.001). Correlation was highest with the AIF(AIL) with a 4-second delay [ρ=0.90, 95% confidence interval (CI): 0.80-0.95] and lowest with the AIF(NLA) (ρ=0.49, 95% CI: 0.11-0.68). Bland-Altman analysis showed the greatest bias for the AIF(NLA) (-42.14±55.60 mL) and the smallest bias for the AIF(AIL) with 4-second delayed (5.18±29.35 mL). Spatial agreement was also best with the AIF(AIL) with a 4-second delay (median Dice coefficient 0.55) and poorest with the 8-second input (0.43). CONCLUSIONS: It is feasible to automatically select AIF(AIL) derived from lesions to improve the accuracy of hypoperfused tissues in CTP. This approach may reduce overtreatment and support more precise diagnosis and management of ischemic stroke.