Synthesizing High-<i>b</i>-Value Diffusion-weighted Imaging of the Prostate Using Generative Adversarial Networks.
retrospective_cohort · Level III
Where this comes from
- Record sourced from PubMed, PMID 34617025.
- Also identified by DOI 10.1148/ryai.2021200237 and PMC identifier 8489442.
- No licence information is recorded for this record.
- Because redistribution is not established, this page shows the abstract only. Follow the links below for the full text.
Abstract
To develop and evaluate a diffusion-weighted imaging (DWI) deep learning framework based on the generative adversarial network (GAN) to generate synthetic high-<i>b</i>-value (<i>b</i> =1500 sec/mm<sup>2</sup>) DWI (SYN<sub>b1500</sub>) sets from acquired standard-<i>b</i>-value (<i>b</i> = 800 sec/mm<sup>2</sup>) DWI (ACQ<sub>b800</sub>) and acquired standard-<i>b</i>-value (<i>b</i> = 1000 sec/mm<sup>2</sup>) DWI (ACQ<sub>b1000</sub>) sets. This retrospective multicenter study included 395 patients who underwent prostate multiparametric MRI. This cohort was split into internal training (96 patients) and external testing (299 patients) datasets. To create SYN<sub>b1500</sub> sets from ACQ<sub>b800</sub> and ACQ<sub>b1000</sub> sets, a deep learning model based on GAN (M<sub>0</sub>) was developed by using the internal dataset. M<sub>0</sub> was trained and compared with a conventional model based on the cycle GAN (M<sub>cyc</sub>). M<sub>0</sub> was further optimized by using denoising and edge-enhancement techniques (optimized version of the M<sub>0</sub> [Opt-M<sub>0</sub>]). The SYN<sub>b1500</sub> sets were synthesized by using the M<sub>0</sub> and the Opt-M<sub>0</sub> were synthesized by using ACQ<sub>b800</sub> and ACQ<sub>b1000</sub> sets from the external testing dataset. For comparison, traditional calculated (<i>b</i> =1500 sec/mm<sup>2</sup>) DWI (CAL<sub>b1500</sub>) sets were also obtained. Reader ratings for image quality and prostate cancer detection were performed on the acquired high-<i>b</i>-value (<i>b</i> = 1500 sec/mm<sup>2</sup>) DWI (ACQ<sub>b1500</sub>), CAL<sub>b1500,</sub> and SYN<sub>b1500</sub> sets and the SYN<sub>b1500</sub> set generated by the Opt-M<sub>0</sub> (Opt-SYN<sub>b1500</sub>). Wilcoxon signed rank tests were used to compare the readers' scores. A multiple-reader multiple-case receiver operating characteristic curve was used to compare the diagnostic utility of each DWI set. When compared with the M<sub>cyc</sub>, the M<sub>0</sub> yielded a lower mean squared difference and higher mean scores for the peak signal-to-noise ratio, structural similarity, and feature similarity (<i>P</i> < .001 for all). Opt-SYN<sub>b1500</sub> resulted in significantly better image quality (<i>P</i> ≤ .001 for all) and a higher mean area under the curve than ACQ<sub>b1500</sub> and CAL<sub>b1500</sub> (<i>P</i> ≤ .042 for all). A deep learning framework based on GAN is a promising method to synthesize realistic high-<i>b</i>-value DWI sets with good image quality and accuracy in prostate cancer detection.<b>Keywords:</b> Prostate Cancer, Abdomen/GI, Diffusion-weighted Imaging, Deep Learning Framework, High <i>b</i> Value, Generative Adversarial Networks© RSNA, 2021 <i>Supplemental material is available for this article.</i>