Performance of Deep Learning Reconstruction for Detection of Early Ischemic Changes in NCCT: Comparison with ASIR-V in Acute Stroke.

PURPOSE: Evaluation of the impact of Deep Learning Image Reconstruction (DLIR) compared to Adaptive Statistical Iterative Reconstruction-Veo (ASIR-V) on image quality and early ischemic changes detection on non-contrast computed tomography (NCCT) in stroke suspected patients. A secondary objective was to determine the potential influence of reconstruction algorithm on ASPECT scoring relative to automated e-ASPECT score. METHODS: Consecutive patients undergoing NCCT within 6 hours of symptom onset were retrospectively included. Images were reconstructed using ASIR-V and high-strength DLIR. Four readers with varying experience independently assessed subjective image quality, gray-white matter differentiation, diagnostic confidence, presence of ischemic lesions and ASPECTS scoring. Diagnostic performance (accuracy, sensitivity, specificity) was calculated using e-ASPECTS as reference. Evaluation time was recorded. RESULTS: DLIR significantly improved subjective image quality and gray-white matter contrast compared with ASIR-V (odds ratios 2.96-3.96; p 0.001). Diagnostic performance for detecting early ischemic changes showed no significant difference, with similar accuracy, sensitivity and specificity. Evaluation time did not differ. A trend toward higher specificity and reduced bias for ASPECTS ≥6 was observed with DLIR, but mixed-model analysis did not confirm statistical significance. CONCLUSION: DLIR improves subjective image quality in acute stroke NCCT but does not significantly improve detection accuracy or ASPECTS scoring compared with ASIR-V. A tendency toward improved specificity was observed; further studies with larger cohorts are needed.