Vendor-neutral deep learning reconstruction of dynamic contrast-enhanced prostate MRI: Image quality improvement and preservation of diagnostic performance.

PURPOSE: To investigate the effects of vendor-neutral image-domain deep learning reconstruction (DLR) applied retrospectively to dynamic contrast-enhanced (DCE) images from prostate multiparametric magnetic resonance imaging (MRI) on image quality and diagnostic performance for clinically significant prostate cancer (csPCa).

MATERIALS AND

METHODS

This retrospective study included 173 patients who underwent prebiopsy 3T prostate MRI from January 2016 to December 2022 before biopsy. Two radiologists independently assessed qualitative image quality and assigned Prostate Imaging Reporting and Data System (PI-RADS) version 2.1 scores in blinded sessions using multiparametric MRI (mpMRI) sets with original or DLR DCE images. PI-RADS 3+1 denoted DCE-driven upgrading of peripheral zone diffusion-weighted imaging category 3 lesions. Diagnostic performance for csPCa, defined as grade group ≥ 2, was evaluated using receiver operating characteristic curve analysis with area under the curve (AUC) and binary metrics.

RESULTS

DLR DCE images showed improved image noise, capsule sharpness, lesion conspicuity, and confidence in assessing DCE positivity (all P < 0.01) for both readers. Image texture naturalness was lower with DLR DCE images (P < 0.001). The AUC values did not differ significantly between the two mpMRI sets. At the PI-RADS ≥ 4 threshold, Reader 2 showed higher sensitivity with DLR DCE images than with original DCE images (88.5 % vs. 75.4 %, P = 0.027). DLR DCE images yielded fewer PI-RADS 3+1 and more PI-RADS 4-5 assessments. Interreader agreement for PI-RADS categorization and DCE positivity was greater with DLR DCE images.

CONCLUSION

Vendor-neutral image-domain DLR improved the qualitative image quality of prostate DCE images while preserving PI-RADS-based diagnostic performance for csPCa.