Feasibility of deep learning-based super-resolution reconstruction for endometrial cancer T2-weighted imaging

Background No study has assessed the application of deep learning (DL) networks in combination with more advanced super-resolution (SR) reconstruction for endometrial cancer (EC) MRI and the influence of this methods on myometrial invasion. We aimed to evaluate the performance and clinical value of DL-based SR reconstruction for image quality improvement in under-sampling accelerated EC T2WI. Methods Two sets of T2WI were acquired for 23 volunteers with low (T2 _LR ) and high (T2 _HR ) resolution, respectively. For 41 EC patients, T2 _LR and MultiVane T2WI (T2 _MV ) were performed. T2 _SR images were reconstructed from T2 _LR through DL network. Image quality and diagnostic accuracy were evaluated by radiologists. Results For 23 volunteers, the signal-to-noise ratio and peak signal-to-noise ratio of T2 _SR were notably higher than those of T2 _LR and T2 _HR , and the edge rise distance of T2 _SR was lower ( P  < 0.005). For 41 EC patients, the image sharpness, capsule delineation, overall image quality, and diagnostic confidence of T2 _SR and T2 _MV were notably higher than those of T2 _LR ( P  < 0.005). Compared with T2 _MV , the scanning time of T2 _SR was reduced by 61% (50s vs. 128s). T2 _SR and T2 _MV images had similar diagnostic accuracy for myometrial invasion assessed by radiologists. Conclusions DL-based super-resolution reconstruction significantly improved the image quality of uterus T2WI compared to the conventional iterative reconstruction by compressed sensing, saving scanning time and ensuring the accurate evaluation of myometrial invasion of endometrial cancer.