Deep Learning Reconstruction Versus Hybrid Iterative Reconstruction for Acute Cerebral Infarction Detection on 135 kVp Non-Contrast Brain CT.

Abstract

Purpose: Deep learning reconstruction (DLR) is useful to reduce image noise and improve contrast resolution compared with hybrid iterative reconstruction (Hybrid IR). This study compared image quality and infarct detection between DLR and Hybrid IR using thin-slice brain CT.

Materials and methods: Eighty-one patients (39 with acute infarction, 42 without) underwent 135 kVp non-contrast brain CT and MRI within 24 h of admission. CT images (2-mm thickness) were reconstructed using both Hybrid IR and DLR. Image noise was measured in white matter, gray matter, and infarct lesions. Three general radiologists independently assessed infarct presence. Sensitivity was evaluated using patient- and region-based analyses.

Results: DLR demonstrated significantly lower image noise than Hybrid IR in white matter, gray matter, and infarct lesions (1.69 vs. 4.40, 1.43 vs. 3.93, and 1.68 vs. 3.94 HU, respectively; all p < 0.001). Contrast-to-noise ratio was significantly higher with DLR (5.10 vs. 2.36, p < 0.001). In patient-based analysis, infarct detection sensitivity was higher with DLR (66.7%-71.8%) than with Hybrid IR (59.0%-69.2%) (p > 0.05). In region-based analysis, DLR showed significantly higher sensitivity for one reader (60.5% vs. 50.0%, p = 0.004).

Conclusion: In this study, DLR significantly reduces image noise and improves contrast-to-noise ratio in thin-slice brain CT. These improvements may help general radiologists in diagnosing acute cerebral infarction.