Comparing prostate diffusion weighted images reconstructed with a commercial deep-learning product to a deep learning phase corrected model at 1.5 T

Abstract

Purpose: To determine whether a new deep learning (DL) based phase corrected (DLPC) reconstruction model can enhance image quality of diffusion weighted images of the prostate acquired at 1.5 T compared to a commercially available DL based product.

Methods and materials: A retrospective study of 30 consecutive patients undergoing conventional multiparametric MRI (mpMRI) of the prostate on a single 1.5 T scanner was performed. Diffusion image datasets reconstructed with a commercially available DL product and a new DLPC model were assessed. Qualitative image assessment was performed by three board certified radiologists using a 5-point Likert scale across four features and inter-rater agreement was estimated using Gwet's AC2 statistic. Quantitative image comparison was performed by assessing SNR of acquired intermediate b-value (b = 1000 s/mm²) diffusion images. The Wilcoxon matched-pairs signed rank test was used to assess differences between techniques. Image noise was assessed using the edge function.

Results: Median patient age was 70 years (interquartile range: 66.0-75.3). All radiologists perceived less noise and better image quality for all DLPC image sets compared to commercial DL images (p < 0.05). Significantly higher SNR was observed for the acquired intermediate b-value diffusion images reconstructed with DLPC (median SNR: 49.4 vs 27.5; p < 0.001), and mean ADC values did not significantly differ between DLPC and DL images (p = 0.63). Edge analyses demonstrated significantly reduced noise for DLPC images (p < 0.001).

Conclusions: DLPC image reconstruction of diffusion weighted prostate image datasets reduces image noise and improves SNR over a commercial DL product at 1.5 T.

Keywords: Artificial intelligence (AI); Deep learning (DL); Diffusion weighted imaging (DWI); Magnetic resonance imaging (MRI); Prostate cancer.