Multiparametric AI-based perfusion analysis outperforms Tmax thresholding for critically hypoperfused tissue estimation in acute ischemic stroke undergoing admission MRI.

OBJECTIVE: Perfusion imaging is used to estimate critically hypoperfused tissue in acute ischemic stroke (AIS), commonly using threshold-based methods. This study compared an artificial intelligence (AI)-driven multiparametric approach with conventional thresholding for estimating ischemic core and hypoperfused tissue volumes. MATERIALS AND METHODS: We analyzed 186 AIS patients from the HIBISCUS-STROKE cohort (NCT03149705) who underwent MRI before mechanical thrombectomy. A deep-learning AI model, incorporating multiparametric imaging features, was compared to standard thresholding methods. The ischemic core was defined by an ADC ≤ 620 × 10⁻⁶mm2/s threshold, and critically hypoperfused tissue was defined by Tmax ≥ 6 s. Successful thrombectomy was determined by a modified thrombolysis in cerebral infarction score ≥ 2B. For recanalized patients, AI and ADC-based core volumes were compared with the final infarct volume from day 6 T2-FLAIR MRI. For non-recanalized patients, the hypoperfused tissue estimates were compared with the final infarct volume. RESULTS: The AI-derived ischemic core volume strongly correlated with ADC-based estimates (ρ = 0.82, P < 0.0001) with no significant volume differences (P = 0.38). In recanalized patients, both methods correlated similarly with final infarct volume (ρ≈0.67; P < 0.0001), with comparable median underestimations and biases on Bland-Altman analysis. AI-estimated hypoperfused tissue volumes were significantly lower than those obtained via Tmax thresholding (medians: 44.4 mL vs. 79.6 mL; P = 0.007), with a more favorable bias in non-recanalized patients. CONCLUSION: The AI-driven multiparametric approach refined the estimation of critically hypoperfused tissue compared to conventional Tmax thresholding while maintaining comparable performance for assessing the ischemic core.