Image Quality and Diagnostic Performance of Accelerated T2-weighted Imaging of Prostate with Deep Learning Reconstruction: A Comparative Study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Kobe University Graduate School of Medicine, Kobe Hyogo, Japan.
- Department of Diagnostic Radiology, Kindai University Faculty of Medicine, Sakai Osaka, Japan.
- Division of Radiological Technology, Department of Radiology, Kobe University Hospital, Kobe Hyogo, Japan.
- Department of Urology, Kobe University Graduate School of Medicine, Kobe Hyogo, Japan.
Abstract
To compare accelerated T2-weighted turbo spin-echo imaging with deep learning reconstruction (DLR-TSE) with conventional T2-weighted TSE (conv-TSE) and accelerated TSE without DLR (non-DLR-TSE), and to evaluate image quality and diagnostic performance of Prostate Imaging Reporting and Data System version 2.1 (PI-RADS v2.1)-based T2 scoring in prostate MRI. This single-center retrospective study included 60 patients who underwent prostate MRI with all 3 T2-weighted image sets acquired in the same examination. Qualitative image quality was independently assessed by 2 radiologists using 6 parameters on a 5-point Likert scale. Quantitative metrics included apparent SNR (aSNR), apparent contrast-to-noise ratio (aCNR), and contrast ratio (CR). Diagnostic performance of PI-RADS T2 scores for transition zone lesions was evaluated using receiver operating characteristic (ROC) analysis, sensitivity, specificity, and accuracy. Noninferiority of DLR-TSE relative to conv-TSE was tested with predefined margins. Inter-reader agreement was assessed using weighted kappa statistics. DLR-TSE demonstrated noninferiority to conv-TSE for all qualitative parameters and quantitative metrics for both readers. Both DLR-TSE and conv-TSE showed significantly higher image quality scores and quantitative values than non-DLR-TSE. For PI-RADS T2 scoring, DLR-TSE achieved diagnostic performance comparable to conv-TSE. For reader 1, the area under the ROC curve (AUC) was identical for DLR-TSE and conv-TSE (0.83; 95%CI 0.69-0.95), and significantly higher than for non-DLR-TSE (0.70; 95%CI 0.56-0.83). Specificity and overall accuracy were markedly reduced with non-DLR-TSE for both readers, whereas sensitivity did not differ significantly among methods. Inter-reader agreement was substantial to almost perfect for DLR-TSE and conv-TSE, and lower for non-DLR-TSE. DLR-TSE reduced acquisition time by approximately 60% compared with conv-TSE. Accelerated T2-weighted imaging with DLR allows substantial reduction in scan time while maintaining image quality, diagnostic performance, and inter-reader agreement comparable to those of conventional T2-weighted imaging. DLR-TSE may serve as a practical option for improving examination efficiency in clinical prostate MRI.