Deep-learning reconstruction for noncontrast head computed tomography: Improved image quality and potential diagnostic impact in acute ischemic stroke

Deep learning reconstruction is a promising technique for improving head CT image quality compared to traditional methods. The purpose of this study was to compare image quality and the potential diagnostic impact of vendor-agnostic deep-learning reconstruction (DLR) versus hybrid iterative reconstruction (IR) for the evaluation of acute ischemic stroke on noncontrast head computed tomography (CT). This single-institution retrospective study included 100 patients (mean age 69.6 years; 44 women) who visited the emergency department with suspected acute stroke and underwent noncontrast head CT between November 2021 and November 2022. Fifty patients were confirmed to have acute infarction lesions via subsequent DWI, while the remaining 50 had negative DWI results. The CT images were reconstructed with both IR and DLR. Four reviewers (two radiology residents and two neuroradiologists) assessed subjective image quality and the conspicuity of acute ischemic lesions on a 5-point scale. Nonparametric comparisons of scores and ROC curve analysis were performed. Compared with IR, DLR resulted in superior overall image quality (4.02 vs. 3.23, p < 0.01). While no significant differences were found in lesion conspicuity according to one neuroradiologist (3.90 vs. 3.70; p = 0.14), other experienced readers reported significantly enhanced conspicuity with DLR(3.60 vs. 3.12, p = 0.014; 3.84 vs. 3.12, p < 0.01; and 4.06 vs. 3.62, p = 0.018, respectively). ROC curve analysis revealed an increase in the area under the curve (AUC) for radiology resident readers (0.888 vs. 0.844, p = 0.002) and a decrease in AUC for neuroradiologists (0.857 vs. 0.911, p = 0.034). Our study demonstrates that DLR offers subjective improvements in image quality, tissue differentiation, and noise reduction. We observed improvement in the diagnostic performance of less experienced radiologists with DLR .