Improved image quality in the Beat Mapping using a Convolutional Recurrent Neural Network (BeatMapCRNN)

Motivation: Cardiovascular magnetic resonance parametric mappings have high diagnostic and prognostic value. However, the image quality is significantly degraded by noise, especially saturation-based T1 mapping. A new technique for image quality enhancement is required. Goal(s): To develop a neural network to enhance image quality in the most commonly used cardiac parametric mapping sequences by removing noise. Approach: A convolutional recurrent neural network (BeatMapCRNN) with local and non-local mean convolutional blocks was developed and tested by mapping data of healthy volunteers and patients by MOLLI, SASHA, and T2-prep bSSFP mapping. Results: BeatMapCRNN could effectively remove the noise and improve map quality. Impact: This study developed a convolutional recurrent neural network to alleviate noise artifacts in cardiac T1 and T2 mapping. Validation indicated that the proposed method could improve map quality for most used T1/T2 sequences without compromising accuracy or blurring lesions.