CROSS-VENDOR PERFORMANCE OF POST-ACQUISITION DICOM-BASED DEEP LEARNING RECONSTRUCTION IN ACCELERATED PROSTATE MRI: A PROSPECTIVE MULTI-CENTER STUDY COMPARING SCANNER-INTEGRATED K-SPACE RECONSTRUCTION

Abstract

Deep learning reconstruction (DLR) has substantially improved image quality in accelerated magnetic resonance imaging (MRI); however, most commercially available approaches rely on vendor-specific k-space reconstruction integrated into proprietary scanner platforms, limiting interoperability across heterogeneous imaging environments. This prospective multicenter study evaluated whether a vendor-agnostic, post-acquisition DICOM-based DLR algorithm could achieve image quality and diagnostic performance comparable to scanner-integrated k-space DLR for accelerated multiparametric prostate MRI. Total 120 men (with average age, 67.4 ± 9.2 yrs) underwent standard and accelerated MRI acquisitions (R=2 and R=4) on scanners from three major manufacturers. Accelerated datasets were reconstructed using either a vendor-agnostic DICOM-based DLR algorithm or each manufacturer's native k-space DLR solution. At an acceleration factor of R=4, vendor-agnostic DLR demonstrated non-inferior overall image quality (4.1 ± 0.7 vs. 4.3 ± 0.6; p = 0.42), with comparable Signal Strength-to-Noise Ratio, Contrast Enhancement-to-noise ratio, structural similarity Metric, peak signal-quality ratio, diagnostic confidence, and PI-RADS agreement (κ=0.82 vs. 0.84). The vendor-agnostic approach achieved significantly greater scan-time reduction (61.3% vs. 52.3%; p<0.01), with consistent performance across all scanner vendors. These findings demonstrate that vendor-agnostic DICOM-based DLR enables platform-independent acceleration of prostate MRI while maintaining diagnostic image quality, supporting broader clinical deployment without dependence on proprietary reconstruction hardware or software.