Luo Lingfeng, Ye Chen, Li Tianxian, Zhong Ming, Wang Lihui, Zhu Yuemin
Key Laboratory of Advanced Medical Imaging and Intelligent Computing of Guizhou Province, Engineering Research Center of Text Computing & Cognitive Intelligence, Ministry of Education, State Key Laboratory of Public Big Data, College of Computer Science and Technology, Guizhou University, Guiyang, China.
Department of Radiology, International Exemplary Cooperation Base of Precision Imaging for Diagnosis and Treatment, NHC Key Laboratory of Pulmonary Immune-related Diseases, Guizhou Provincial People's Hospital, Guiyang, China.
Med Phys. 2025 Jul;52(7):e17825. doi: 10.1002/mp.17825. Epub 2025 Apr 14.
The intravoxel incoherent motion (IVIM) parameter estimation is affected by noise, while existing CNN-based fitting methods utilize neighborhood spatial features around voxels to obtain more robust parameters. However, due to the heterogeneity of tissue, neighborhood features with low similarity can lead to excessively smooth parameter maps and even loss of tissue details.
To propose a novel neural network fitting approach, IVIM-CNN, which utilizes similar neighborhood information of voxels to assist in the estimation of IVIM parameters in diffusion-weighted imaging (DWI).
The proposed fitting model is based on convolutional neural network (CNN), which first identifies the similar neighborhoods of voxels through cluster analysis and then uses CNN to learn the spatial features of similar neighborhoods to reduce the impact of noise on the parameter estimation of the voxel. To evaluate the performance of the proposed method, comparisons were conducted with the least squares (LSQ), Bayesian, PI-DNN, and IVIM-CNN algorithms on both simulated and in vivo brains, including 23 healthy brains and three brain tumors, in terms of root mean square error (RMSE) of IVIM parameters and the parameter contrast ratio between the tumor and normal regions.
The CNN-based methods, such as IVIM-CNN and IVIM-CNN, yield smoother parameter maps compared to voxel-based methods like nonlinear least squares, segmented nonlinear least squares, Bayesian, and PI-DNN. Additionally, the IVIM-CNN retains more local tissue details while maintaining smoothness of parameter maps compared to the IVIM-CNN. In simulated experiments, IVIM-CNN outperforms IVIM-CNN in terms of parameter estimation accuracy (SNR = 30; RMSE [ ] = 0.0168 vs. 0.0253; RMSE ( ) = 0.0001 vs. 0.0002; RMSE [ ] = 0.0266 vs. 0.0416). In addition, compared with other methods, the proposed IVIM-CNN is more robust to noise, which is reflected in the lower RMSE of each parameter at different SNRs. For in vivo brains, compared to other methods, IVIM-CNN achieved the highest PCR for most parameters when comparing the normal and tumor regions.
The IVIM-CNN method uses similar neighborhood information to assist IVIM parameter fitting by reducing the impact of noise on voxel parameter estimation, thereby improving the accuracy of parameter estimation and increasing the potential for IVIM clinical application.
体素内不相干运动（IVIM）参数估计受噪声影响，而现有的基于卷积神经网络（CNN）的拟合方法利用体素周围的邻域空间特征来获得更稳健的参数。然而，由于组织的异质性，相似度低的邻域特征会导致参数图过度平滑，甚至丢失组织细节。
提出一种新颖的神经网络拟合方法IVIM-CNN，该方法利用体素的相似邻域信息来辅助扩散加权成像（DWI）中IVIM参数的估计。
所提出的拟合模型基于卷积神经网络（CNN），它首先通过聚类分析识别体素的相似邻域，然后使用CNN学习相似邻域的空间特征，以减少噪声对体素参数估计的影响。为了评估所提方法的性能，在模拟脑和活体脑（包括23个健康脑和3个脑肿瘤）上，就IVIM参数的均方根误差（RMSE）以及肿瘤与正常区域之间的参数对比率，将其与最小二乘法（LSQ）、贝叶斯法、PI-DNN和IVIM-CNN算法进行比较。
与基于体素的方法（如非线性最小二乘法、分段非线性最小二乘法、贝叶斯法和PI-DNN）相比，基于CNN的方法（如IVIM-CNN）生成的参数图更平滑。此外，与IVIM-CNN相比，IVIM-CNN在保持参数图平滑的同时保留了更多局部组织细节。在模拟实验中，IVIM-CNN在参数估计准确性方面优于IVIM-CNN（SNR = 30；RMSE[ ] = 0.0168对0.0253；RMSE（ ） =
0.0001对0.0002；RMSE[ ] = 0.0266对0.0416）。此外，与其他方法相比，所提的IVIM-CNN对噪声更具鲁棒性，这体现在不同SNR下每个参数的RMSE更低。对于活体脑，与其他方法相比，在比较正常区域和肿瘤区域时，IVIM-CNN对大多数参数实现了最高的PCR。
IVIM-CNN方法通过减少噪声对体素参数估计的影响，利用相似邻域信息辅助IVIM参数拟合，从而提高参数估计的准确性，并增加IVIM临床应用的潜力。
