基于深度学习的三翻转角2D SPGR MRI序列定量估算单侧肾动脉狭窄动物模型肾脏R1参数图的初步研究  

A Preliminary Study on Quantitative Estimation of Renal R1 Parameter Map in an Animal Model of Unilateral Renal Artery Stenosis Based on Deep Learning of Triple Flip Angle 2D SPGR MRI Sequences

作　　者：米悦[1] 张晓东[2] 吴静云[2] 孙艳[2] 罗健[2] 赵凯[2] 王霄英[2] MI Yue;ZHANG Xiaodong;WU Jingyun(Department of Urology,Peking University First Hospital,Beijing 100034,P.R.China)

机构地区：[1]北京大学第一医院泌尿外科,100034 [2]北京大学第一医院医学影像科,100034

出　　处：《临床放射学杂志》2023年第3期450-454,共5页Journal of Clinical Radiology

基　　金：北京大学第一医院科研种子基金(编号:2021SF45)。

摘　　要：目的探讨基于深度学习实现对三翻转角2D SPGR MRI序列定量估算单侧肾动脉狭窄动物模型肾脏R1参数图的可行性。方法共纳入12只平均体重3.2 kg的健康新西兰大白兔,对每只兔子施行左肾动脉部分结扎手术以建立单侧肾动脉狭窄(RAS)动物模型。RAS术前和术后每隔10 min采集一次2D单层三翻转角SPGR MRI数据,其中术前2次和术后9次,以获得单侧RAS造成肾脏水含量水平改变前后的肾脏R1参数图。最终获得127个2D图像对,每对图像包括一个由相同层面3个翻转角(15°,24°和33°)的SPGR序列图像组成三通道2D图像和一个传统可变翻转角R1估算方法得到相应层面的R1参数图图像。应用基于深度学习的编码器-解码器结构训练R1参数图生成模型。将RAS术后采集的92例数据分为训练集(74对)和调优集(18对),将RAS术前采集的34例数据及1例RAS术后的数据作为测试集(35对)。以测试集的峰值信噪比和结构相似性结果为R1参数图生成模型的评价指标。结果在测试集中,R1参数图生成模型的峰值信噪比(dB)和结构相似性(%)分别为22.08±2.33和79.49±6.49。结论基于深度学习模型可实现对兔子肾脏R1参数图的定量估算进而定量评估其肾实质的水含量。Objective To investigate the feasibility of a deep learning model-based approach for quantitative estimation of renal R1 parameter map in an animal model of unilateral renal artery stenosis using triple flip-angle 2D SPGR MRI sequences.Methods A total of 12 healthy New Zealand rabbits with an average weight of 3.2 kg were enrolled for evaluation,and each rabbit was subjected to partial ligation of the left renal artery to create an animal model of unilateral renal artery stenosis(RAS),and data were collected at 10 minute intervals before and after RAS,including two time points before and nine time points after surgery,for a total of 112D single-layer triple flip angle SPGR MRI data at different time points to obtain R1 parameter maps of the kidney before and after alteration of renal water content levels due to unilateral renal artery stenosis.A total of 1272D image pairs were obtained,each pair consisting of a three-channel 2D image consisting of a sequence of SPGR images with three flip angles(15°,24°and 33°)at the same slice and a conventional variable flip angle R1 estimation method to obtain R1 parametric map images at the corresponding slice.The encoder-decoder structure is used as the base architecture of the R1 parametric map generation model.Ninety-two cases of data collected after RAS surgery were divided into training dataset(74 pairs)and tuning dataset(18 pairs),the data from 34 cases collected before RAS surgery and 1 case after RAS surgery were used as the testing dataset(35 pairs).The peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)results of the test dataset were used as evaluation metrics for the R1 parameter map generation model.Results In the test dataset,the PNSR(dB)and SSIM(%)of the R1 parametric map generation model were 22.08±2.33 and 79.49±6.49,respectively.Conclusion The deep learning model is feasible for quantitative esti-mation of R1 parametric maps of rabbit kidney for triple flip angle SPGR MRI sequences to quantitatively assess the water content of its renal pare

关 键 词：深度学习 人工智能 SPGR序列 肾动脉狭窄 R1参数图 

分 类 号：R445.2[医药卫生—影像医学与核医学] R692[医药卫生—诊断学]