Deep Learning Super Resolution Reconstruction for Fast and Motion Robust T2-weighted Prostate MRI

Overview

The aim of this study was therefore to investigate a new unrolled DL super resolution reconstruction of an initially low-resolution Cartesian T2 turbo spin echo sequence (T2 TSE) and compare it qualitatively and quantitatively to standard high-resolution Cartesian and non-Cartesian T2 TSE sequences in the setting of prostate mpMRI with particular interest in image sharpness, conspicuity of lesions and acquisition time. Furthermore, the investigators assessed the agreement of assigned PI-RADS scores between deep learning super resolution and standard sequences.

Prostate cancer has been among the most prevalent cancer types in men for years, being responsible for 7.8% of all newly diagnosed cases in 2020, holding the 2th place right after lung cancer. Early and non-invasive diagnostics was improved vastly by multiparametric MRI (mpMRI) of the prostate, detecting clinically significant prostate cancer and forming the baseline for guided biopsy of the prostate, while it can prevent unnecessary biopsies in patients with elevated prostate specific antigen, but no visible lesions. With an aging population fast, efficient and highly qualitative MRI scans are needed to satisfy this increasing demand. Deep learning image reconstruction has become increasingly important solving these tasks to produce highly qualitative images while drastically reducing acquisition time. Standard acquisition protocols of prostate mpMRI include T2-weighted, diffusion-weighted and dynamically contrast-enhanced sequences to allow for the classification of prostatic lesions according to the Prostate Imaging Reporting \& Data System (PI-RADS). While the assignment of the PI-RADS score in the peripheral zone of the prostate is mainly determined by the diffusion weighted imaging, the T2-weighted-sequences are mainly responsible for the assessment of the transitional zone. Furthermore, thorough assessment of the prostate necessitates acquisition of T2-weighted sequences in axial and sagittal planes, thus extending acquisition time of MRI protocols. Different approaches have been proposed to accelerate and improve the image acquisition, ranging from the implementation of shortened protocols to the improvement of diffusion weighted sequences or using compressed sensing for the reconstruction of non-Cartesian T2-weighted-sequences. Besides these methods that rely on traditional acquisition and reconstruction methods, deep learning (DL) image reconstruction has become increasingly important solving these tasks to produce highly qualitative images while drastically reducing acquisition time. Despite that, reliable DL-methods for the process of image acquisition and reconstruction of prostate mpMRI itself are sparse. While first approaches for DL denoising has been established, effectively replacing the conventional wavelet function, the remainder of the iterative reconstruction cycle is unaffected and the impact on diagnostic performance of the PI-RADS score remains unclear. Recently developed super resolution deep learning networks are promising to overcome this limitation. First results for DL denoising in different applications, e.g. in musculoskeletal MRI already show good results, leading to significant acceleration of acquisition time while maintaining high image quality. However, the application of these denoising DL-networks in combination with more advanced super resolution networks in prostate mpMRI hasn't been evaluated yet. In this prospective study, between August and November 2022, participants with suspicion for prostate cancer underwent prostate MRI with standard high-resolution Cartesian T2 (T2C) and non-Cartesian T2 (T2NC) sequences. Additionally, a low-resolution Cartesian T2 TSE (T2SR) with DL denoising and super resolution reconstruction was acquired. Artifacts, image sharpness, lesion conspicuity, capsule delineation, overall image quality and diagnostic confidence were rated on a 5-point-Likert-Scale with being non-diagnostic and 5 being excellent. Apparent signal-to-noise ratio (aSNR), contrast-to-noise ratio (aCNR) and edge rise distance (ERD) were calculated. Friedman test and One-way ANOVA were used for group comparisons. Regarding agreement of PI-RADS scores were compared with Cohen's Kappa.