基于深度学习的乳腺DCE-MRI图像推演模型的构建

Construction of the breast DCE-MRI images deduction model based on deep learning

目的:探究利用深度学习算法构建乳腺动态对比增强磁共振成像(dynamic contrast-enhanced magnetic resonance imaging,DCE-MRI)图像推演模型的可行性,寻找无须实际扫描即可获取DCE-MRI图像的新技术。方法:收集2020年1月—2021年8月在西安国际医学中心医院行乳腺MRI检查的54例女性患者的临床信息及影像学资料,随机分为训练集(38例)、测试集1(8例)及测试集2(8例)。基于训练集中患者的T1加权成像(T1-weighted imaging,T1WI)及DCE序列第一期图像,利用深度学习Pix2Pix算法构建乳腺DCE-MRI图像推演模型。在测试集中合成DCE图像,采用峰值信噪比(peak signal to noise ratio,PSNR)、结构相似性(structural similarity,SSIM)、均方误差(mean square error,MSE)、平均绝对误差(mean absolute error,MAE)指标测试模型性能。另外,评估合成的DCE序列与原始序列间的相关性。结果:共合成DCE第一期图像760张。与原始图像相比,合成的图像具有良好的SSIM,较低的信息损失,其中SSIM为0.71±0.01,PSNR为23.70±1.41。另外,合成的DCE序列重建误差较小,MAE为0.032±0.004,MSE为0.006±0.002,且与原始DCE序列呈显著正相关(r=0.872±0.038,95%CI:0.870~0.874,P=0.000)。结论:乳腺DCE-MRI图像推演模型能够自动生成DCE序列第一期图像,为补充非增强MRI序列中信息缺失提供了新思路,同时避免了造影剂的使用、缩短扫描时间,为乳腺MRI筛查的推广应用奠定了基础。

Objective:To investigate the feasibility of building the breast dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)images deduction model by using deep learning algorithms,with the aim of obtaining the DCE-MRI sequences without actual scanning.Methods:Clinical data and MRI images of 54 subjects with breast MRI examinations were collected retrospectively in Xi’an International Medical Center Hospital from January 2020 to August 2021 and divided into a training set(38 subjects),a test set 1(8 subjects)and a test set 2(8 subjects).Here,we built the breast DCE-MRI images deduction model using the Pix2Pix algorithm based on the T1-weighted imaging(T1WI)and the first post-contrast phase of DCE sequence images in the training set.Furthermore,the model performance for synthesizing images and sequences was evaluated in the test set using several metrics such as peak signal to noise ratio(PSNR),structural similarity index measure(SSIM),mean squared error(MSE),and mean absolute error(MAE).In addition,the correlation between the synthesized DCE sequences and the original DCE sequences was evaluated.Results:In total,760 images of the first post-contrast phase of DCE were synthesized using the model.The synthesized images had good structural similarity and low information loss compared with the original images,where the mean SSIM was 0.71±0.01 and PSNR was 23.70±1.41.In addition,we found that the synthesized DCE sequences had small reconstruction errors,MAE was 0.032±0.004,MSE was 0.006±0.002,and the synthesized first post-contrast phase of DCE sequences had a significantly positive correlation with the original sequences,(r=0.872±0.038,95%CI:0.870-0.874,P=0.000).Conclusion:The model which we built can automatically synthesize the first post-contrast phase of DCE sequence images,which provides a new idea for supplementing the missing information in non-enhanced sequences.At the same time,it avoids the application of contrast agents,shortens the scanning time,and lays a solid foundation for the popularization and application of breast MRI screening.