Comparison of deep learning reconstruction and adaptive statistical iterative reconstruction for head CT in acute stroke.

PURPOSE: To compare deep learning reconstruction (DLR) at three denoising strengths (Low, Medium, High) with adaptive statistical iterative reconstruction (ASIR-V) for non-contrast head CT (NCCT) in adult patients presenting with acute stroke symptoms.

METHODS

In this retrospective study, 102 consecutive patients (mean age, 68.65 years) underwent NCCT within 24 h of stroke symptom onset between December 2024 and April 2025. Raw data were reconstructed using ASIR-V and DLR-L, DLR-M, and DLR-H. Two blinded radiologists assessed perceived noise, gray–white matter differentiation (supratentorial, deep, infratentorial), image texture, diagnostic confidence, and beam-hardening artifacts on 5-point Likert scales. Quantitative metrics included signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), posterior fossa artifact index (PFAI), and subcalvarial artifact index (SAI). Alberta Stroke Program Early CT Scores (ASPECTS) were recorded. Qualitative data was analysed using Friedman test with post hoc Wilcoxon signed rank tests (Bonferroni corrected) and quantitative data was analyzed using paired t-tests (Benjamini–Hochberg corrected).

RESULTS

DLR-L and DLR-M achieved significantly higher gray–white matter differentiation scores than ASIR-V on pairwise comparisons (p < 0.05). DLR-H demonstrated the lowest perceived noise and lower texture ratings than ASIR-V (p < 0.05). DLR-H had the highest SNR and CNR in most regions, with large effect sizes in the deep nuclei (d = 0.934) and cerebellum (d = 0.917). All DLR strengths significantly reduced PFAI; SAI was significantly lower with DLR-M and DLR-H. Diagnostic confidence, and ASPECTS did not differ significantly between these techniques.

CONCLUSION

DLR improves image quality and decreases artifacts compared with ASIR-V. Medium denoising strength (DLR-M) provides an optimal balance of artifact reduction and texture preference.