Preview

PDF, English Download (2MB) | Terms of use

Abstract

Magnetic Resonance Imaging (MRI) is widely used in oncology for tumor staging, treatment response assessment, and radiation therapy (RT) planning. However, optimization of MRI sequences for specific clinical needs is complex and very time-consuming due to the large number of parameter settings. This study proposes two different frameworks for the automatic optimization of MRI sequences addressing two clinical use cases in RT planning based on the sequence parameter sets (SPS): I) a regression-based optimization and II) an on-the-run optimization. A phantom with 7 in-house fabricated contrasts was used for measurements. In the regression-based optimization, two prediction models, i) a Generalized Additive Model (GAM), and ii) a Deep Learning (DL) model, were implemented based on a large number of acquired datasets. In contrast, the on-the-run optimization of the SPS was applied directly on the MR scanner using the interface Access-i. Both frameworks used a derivative-free optimization algorithm to iteratively update a parameterized sequence based on the prediction model or on the use of the MR scanner. In each iteration, the mean squared error (MSE) was calculated. Two clinically relevant optimization goals were pursued: achieving the same contrast as in a target image and maximizing the contrast between specified tissue types. Both goals were evaluated using two optimization methods: a covariance matrix adaptation evolution strategy (CMA-ES) and a genetic algorithm (GA). The obtained results demonstrated the potential of the framework for automatic contrast optimization of MRI sequences. Both CMA-ES and GA methods showed promising results in achieving the two optimization goals; however, CMA-ES converged much faster compared to GA. The proposed frameworks enable fast automatic contrast optimization of MRI sequences based on SPS and may be used to enhance the quality of MRI images for dedicated applications in MR-guided RT.