Universal mapping and patient-specific prior implicit neural representation for enhanced high-resolution MRI in MRI-guided radiotherapy.

Abstract

Background: Magnetic resonance imaging (MRI), known for its superior soft tissue contrast, plays a crucial role in radiation therapy (RT). The introduction of MR-LINAC systems enables the use of on-board MRI for adaptive radiotherapy (ART) on the day of treatment to maximize treatment accuracy.

Purpose: Due to patient comfort considerations and the time constraints associated with adaptive radiation therapy (ART), reducing the resolution of on-board MRI to accelerate image acquisition can improve efficiency, especially when acquiring multiple MRIs with different contrast weightings. However, the low-resolution imaging makes it challenging to identify key anatomical structures, potentially limiting treatment precision. To address this challenge, super-resolution of on-board MRI has emerged as a viable solution.

Methods: To achieve super-resolution for on-board MRI, this study proposed a universal anatomical mapping and patient-specific prior implicit neural representation (USINR) framework. Unlike traditional methods that interpolate solely based on individual on-board MR images, USINR can fully utilize the patient-specific anatomical information from a high-resolution prior MRI. In addition, USINR leverages knowledge about universal mapping between population-based prior MRIs and on-board MRIs, elevating the upper bound of super-resolution performance and enabling faster on-board fine-tuning.

Results: USINR was evaluated on three datasets, including IXI, BraTS, and an in-house abdominal dataset. It achieved state-of-the-art performance on all of them. For example, on the BraTS dataset, USINR was trained on 1151 paired training samples (for universal anatomical mapping) and tested on 50 patients. It achieved average SSIM, PSNR, and LPIPS scores of 0.9656, 37.12, and 0.0214, respectively, significantly outperforming the published state-of-the-art method SuperFormer, whose corresponding scores were 0.9488, 35.83, and 0.0388. Furthermore, USINR can complete patient-specific training in less than one minute, rendering it a favorable solution in time-constrained ART workflows. In addition to large-scale dataset evaluations, a case study was conducted on an in-house patient at UT Southwestern Medical Center. This case study included two MRI scans (a prior scan for plan simulation and a new one for on-board imaging) from a single patient with a long interval between two scans, during which the tumor size underwent a significant change. Despite these substantial anatomical changes between prior and on-board imaging, USINR was able to accurately capture the change in tumor size, highlighting its robustness for clinical applications.

Conclusions: By combining knowledge of universal anatomical mapping with patient-specific prior implicit neural representation, USINR offers a novel and reliable approach for MRI super-resolution. This method enhances the spatial resolution of MR images with minimal processing time, thereby balancing the need for image quality and the efficiency of MRI-guided adaptive radiotherapy.