【摘要】 目的：探究无监督深度学习模型将大量临床诊断性结构磁共振扫描重构为高分辨数据的可行性，并应用于脑形态学定量测量。方法：纳入上海市同济医院具有厚层或薄层常规临床结构磁扫描的被试的影像学及临床学数据共580例，包括363例兼具厚层和薄层扫描的受试以及217名仅有薄层扫描的受试。本研究提出一个半监督生成对抗网络框架用于MRI超分辨重建。评估了超分辨图像在各种形态定量分析场景中的实用性。结果：所提方法明显提高了超分辨图像的质量，峰值信噪比超过30 dB，结构相似度达到0.88，平均绝对误差和感知相似度低至0.03。超分辨图像形态学定量值与真实值存在较高一致性。结论：本研究证明了基于生成对抗网络的超分辨算法辅助临床诊断数据用于疾病定量化分析的可靠性，本研究有望回收大量常规诊断性MRI，并将其转化为可用于神经测量学研究的有用数据。更多还原

【Abstract】 Objective: To investigate the viability of using and unsupervised deep learning model to reconstruct a large number of clinical diagnostic structural magnetic resonance scans into high-resolution data, and apply it to brain morphology evaluation. Methods:Imaging and clinical data of 580 subjects with thick-slice or thin-slice structure magnetic scanning were included from Shanghai Tongji Hospital,including 363 subjects with both thick-slice and thin-slice scans and 217 subjects with solely thin-slice images. In this study, a semi-supervised generative adversarial network framework is proposed for MRI super-resolution reconstruction. The practicability of super-resoluted images in various morphological quantitative analysis scenarios is evaluated. Results: The proposed method significantly improved the quality of superresoluted images, the peak signal-to-noise ratio(PSNR) exceeded 30dB, the structural similarity(SSIM) reached 0.88, and the mean absolute error(MAE) and learned perceptual image patch similarity(LPIPS) were lower than 0.03. Conclusion: This study demonstrates the reliability of the super-resolution algorithm based on generative adversarial network to assist clinical diagnostic data for the quantitative analysis of disease,and this study is expected to recover a large number of routine diagnostic MRI and transform it into useful data that can be used in neuromeric research.更多还原