DSFNet: Deep Fusion Parallel MRI Reconstruction Based on SPIRiT and Complex Convolution

Abstract

Long data acquisition time is an inherent disadvantage of magnetic resonance imaging (MRI). To accelerate the data acquisition speed of MRI, undersampling is required, which results in low imaging quality. Based on an iterative self-consistent parallel imaging reconstruction (SPIRiT) and improved complex convolutional neural networks, a deep SPIRiT fusion network (DSFNet) is proposed to improve the quality of reconstructed images. The DSFNet model first uses the SPIRiT model for the reconstruction of under-sampled k-space data. Subsequently, a cascaded complex convolutional neural network with dense connections is utilized for further reconstruction of the calibrated k-space data. Besides, a data consistency layer ensures the fidelity of the reconstructed image in both k-space and image domains. The final magnitude image is fused from the two parts of the reconstruction magnitude images by the sequentialized model-based Bayesian optimization fusion module at a certain ratio. Experimental results on different knee datasets show that DSFNet can bring about a substantial improvement in the visualization quality, peak signal-to-noise ratio and structural similarity of the reconstructed images compared to the SPIRiT, Deepcomplex and DONet models.

摘要

数据采集时间长是磁共振成像的固有缺点。为了加快磁共振成像的数据采集速度,就需要进行欠采样,而欠采样会造成成像质量低下。为提高重建图像的质量,基于迭代自一致性并行成像重建(SPIRiT)和复数卷积神经网络提出了一种深度SPIRiT融合网络(DSFNet)。具体来说,DSFNet模型首先将SPIRiT算法用于对欠采样的k空间数据进行重建,再将具有密集连接的级联复数卷积神经网络用于对校准后的k空间数据进行进一步重建;在k空间和图像域均使用数据一致性层对图像进行保真,最后将两部分重建得到的幅度图像通过基于序列化模型的贝叶斯优化融合模块按一定比例进行融合。在不同的膝盖成像数据集及不同的欠采样模式下进行实验验证。实验结果表明:DSFNet模型相比于SPIRiT、Deepcomplex和DONet模型,在视觉效果、峰值信噪比和结构相似度方面均有明显提高。