Accelerating MR Thermometry by Complex Fully Convolutional Network

Real-time temperature monitoring of patients using MRI with the PRF method is essential for ensuring the effectiveness and safety of HIFU treatment. However, the process of acquiring MRI data is slow which makes real-time temperature monitoring seemingly impossible. Deep learning methods have recently emerged in the field of accelerating MRI image reconstruction due to their outstanding performance. However, most existing algorithms disregard the phase information of the complex-valued data in the original MRI k-space, as they predominantly employ real-valued operations, like convolutions. Consequently, this paper introduces a Complex UNet network based on the traditional UNet, incorporating a range of complex-valued operations in the complex domain, such as complex convolutions, complex batch normalization, and complex activation functions. Furthermore, the paper compares its performance in temperature measurement tasks with the ZF method and the Real UNet network that operates in the real-valued domain. Experimental results demonstrate that Complex UNet achieves superior accuracy in reconstructing temperature maps while utilizing considerably fewer parameters than Real UNet. This indicates that complex convolutional networks have the potential to significantly enhance the speed and accuracy of temperature measurement, making them highly suitable for real-time temperature monitoring in clinical applications.