Clinical feasibility of a deep learning approach for conventional and synthetic diffusion-weighted imaging in breast cancer: Qualitative and quantitative analyses

Abstract

Purpose: In this study, we aimed to investigate the clinical feasibility of deep learning (DL)-based reconstruction applied to conventional diffusion-weighted imaging (cDWI) and synthetic diffusion-weighted imaging (sDWI) by comparing the DL reconstructions to cDWIs and sDWIs in patients with various breast malignancies. Methods: We retrospectively analyzed 115 patients with biopsy-proven breast malignancies who underwent breast magnetic resonance imaging from July 2022 to June 2023, including cDWI with b-value of 800 s/mm2 (cDWI800), sDWI with b-value of 1500 s/mm2 (sDWI1500), DWI using DL-based reconstruction (DL-DWI) with b-value of 800 s/mm2, and synthetic DL-DWI with b-value of 1500 s/mm2 (DL-DWI800 and sDL-DWI1500). Two radiologists independently performed the qualitative analyses using a 5-point Likert scale for all DWI sets. The quantitative analyses were also conducted for signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and cancer-to-parenchyma contrast ratio (CPR). Results: DL-DWI800 and sDL-DWI1500 provided better lesion conspicuity and thoracic muscle and rib delineation than cDWI800 and sDWI1500 (all P < 0.05). DL-DWI800 and sDL-DWI1500 showed comparable normal parenchymal signals to those of cDWI800 and sDWI1500 (all P > 0.05). sDL-DWI1500 and sDWI1500 showed no significant differences in SNR and CNR (P = 0.908, and P = 0.081, respectively). DL-DWI800 and cDWI800 were not significantly different between SNR, CNR, and CPR (all P > 0.05). Conclusions: DL-DWI outperformed cDWI and sDWI in both qualitative and quantitative analyses at the high b-values, while also achieving a shorter acquisition time. © 2024 Elsevier B.V.