Zou Jiaren, Balter James M, Cao Yue
Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, 48109, USA.
Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI, 48109, USA.
Med Phys. 2020 Aug;47(8):3447-3457. doi: 10.1002/mp.14222. Epub 2020 Jun 3.
T -weighted dynamic contrast-enhanced Magnetic Resonance Imaging (DCE-MRI) is typically quantified by least squares (LS) fitting to a pharmacokinetic (PK) model to yield parameters of microvasculature and perfusion in normal and disease tissues. Such fitting is both time-consuming as well as subject to inaccuracy and instability in parameter estimates. Here, we propose a novel neural network approach to estimate the PK parameters by extracting long and short time-dependent features in DCE-MRI.
A Long Short-Term Memory (LSTM) network, widely used for processing sequence data, was employed to map DCE-MRI time-series accompanied with an arterial input function to parameters of the extended Tofts model. Head and neck DCE-MRI from 103 patients were used for training and testing the LSTM model. Arterial input functions (AIFs) from 78 patients were used to generate synthetic DCE-MRI time-series for training, during which data augmentation was used to overcome the limited size of in vivo data. The model was tested on independent synthesized DCE data using AIFs from 25 patients. The LSTM performance was optimized for the numbers of layers and hidden state features. The performance of the LSTM was tested for different temporal resolution, total acquisition time, and contrast-to-noise ratio (CNR), and compared to the conventional LS fitting and a CNN-based method.
Compared to LS fitting, the LSTM model had comparable accuracy in PK parameter estimations from fully temporal-sampled DCE-MRI data (~3 s per frame), but much better accuracy for the data with temporally subsampling (4s or greater per frame), total acquisition time truncation by 48%-16%, or low CNR (5 and 10). The LSTM reduced normalized root mean squared error by 40.4%, 46.9%, and 53.0% for sampling intervals of 4s, 5s, and 6s, respectively, compared to LS fitting. Compared to the CNN model, the LSTM model reduced the error in the parameter estimates up to 55.2%. Also, the LSTM improved the inference time by ~ 14 times on CPU compared to LS fitting.
Our study suggests that the LSTM model could achieve improved robustness and computation speed for PK parameter estimation compared to LS fitting and the CNN based network, particularly for suboptimal data.
T加权动态对比增强磁共振成像（DCE-MRI）通常通过对药代动力学（PK）模型进行最小二乘（LS）拟合来量化，以得出正常组织和疾病组织中微血管和灌注的参数。这种拟合既耗时，参数估计又容易出现不准确和不稳定的情况。在此，我们提出一种新颖的神经网络方法，通过提取DCE-MRI中长时间和短时间依赖的特征来估计PK参数。
使用广泛用于处理序列数据的长短期记忆（LSTM）网络，将伴随动脉输入函数的DCE-MRI时间序列映射到扩展Tofts模型的参数。来自103名患者的头颈部DCE-MRI用于训练和测试LSTM模型。来自78名患者的动脉输入函数（AIF）用于生成合成DCE-MRI时间序列进行训练，在此期间使用数据增强来克服体内数据量有限的问题。该模型使用来自25名患者的AIF在独立合成的DCE数据上进行测试。针对层数和隐藏状态特征的数量对LSTM性能进行了优化。测试了LSTM在不同时间分辨率、总采集时间和对比噪声比（CNR）下的性能，并与传统的LS拟合和基于卷积神经网络（CNN）的方法进行了比较。
与LS拟合相比，LSTM模型在从全时间采样的DCE-MRI数据（每帧约3秒）估计PK参数时具有相当的准确性，但对于时间下采样（每帧4秒或更长时间）、总采集时间截断48%-16%或低CNR（5和10）的数据，其准确性要高得多。与LS拟合相比，对于采样间隔为4秒、5秒和6秒的情况，LSTM分别将归一化均方根误差降低了40.4%、46.9%和53.0%。与CNN模型相比，LSTM模型将参数估计中的误差降低了高达55.2%。此外，与LS拟合相比，LSTM在CPU上的推理时间提高了约14倍。
我们的研究表明，与LS拟合和基于CNN的网络相比，LSTM模型在PK参数估计方面可以实现更高的鲁棒性和计算速度，特别是对于次优数据。
