X. Zhou1,2, N. Huang2, Z. Fang1,2, P. Zou1,2, K. Yuan3, M. Liu3, J. Lang1,2, M. Chen1,2, and Y. Zhao4; 1School of Medicine, University of Electronic Science and Technology of China, Chengdu, Sichuan, China, 2Department of Radiation Oncology, Radiation Oncology Key Laboratory of Sichuan Province, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, Affiliated Cancer Hospital of University of Electronic Scienc, Chengdu, Sichuan, China, 3Sichuan Cancer Hospital & Institute, University of Electronic Science and Technology of China, Chengdu, Sichuan, China, 4Department of Radiation Oncology, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center; Cancer Hospital affiliate to University of Electronic Science and Technology of China, Chengdu, China
Purpose/Objective(s): The Unity MR-Linac has the potential to allow treatment adaptation using quantitative MRI. Intravoxel incoherent motion (IVIM) can non-invasively reveal information about tissue cells and microcirculation without the need for contrast agents, making it a valuable tool for daily adaptation. However, the accurate measurement of IVIM parameters on the Unity MR-Linac remains a challenge. This study aims to give recommendations on the choice of b-value and fitting methods to accurately measure the IVIM parameter on the Unity MR-Linac for the head and neck. Materials/
Methods: Eleven healthy volunteers underwent a single imaging session of the head and neck on Unity 1.5 Tesla MR-Linac. The session included two repetitions of high NSA (Number of Signals Averaged) DWI and multi-b-value DWI. Images were co-registered using a sub-pixel rigid registration algorithm. The ROIs were delineated in T2 images and propagated to DWI images. The signal-to-noise ratio (SNR) was measured on one volunteer. Twenty thousand simulated diffusion signals were generated by adding Rician noise. For high NSA DWI, the IVIM parameters (diffusion coefficient Dt, perfusion fraction Fp, and pseudo-diffusion coefficient Dp) were calculated using a segmented fitting method. Parallelly, for multi-b-value DWI, Levenberg-Marquardt (LM) regression, polynomial Bayesian estimation, and an unsupervised deep learning approach were used to derive IVIM parameters. Three volunteers were used as training data. The bias of the derived parameter was calculated in simulation. The repeatability of each parameter for the parotid glands was assessed using within-subject coefficients of variation (wCV) in the remaining eight volunteers. Results: In the simulation, for the segmented fitting method, LM regression, polynomial Bayesian estimation and unsupervised deep learning approach, the biases of Dt/Dp/Fp were 0.0007×10-3 mm2/s / 180.10×10-3 mm2/s /19%, 0.06×10-3 mm2/s / 12.58×10-3 mm2/s /1.70%, 0.06×10-3 mm2/s / 12.70×10-3 mm2/s /1.73%, and 0.0008×10-3 mm2/s / 1.33×10-3 mm2/s /0.65%, respectively. In vivo, for the segmented fitting method, LM regression, polynomial Bayesian estimation and unsupervised deep learning approach, the repeatability wCVs of Dt/Dp/Fp were 5.53%/14.01%/7.60%, 9.10%/9.80%/4.86%, 29.23%/12.00%/4.47%, and 7.59%/2.44%/5.54%, respectively. The Dp derived from high NSA DWI using the segmented fitting method presented a huge overestimation. For multi-b-value DWI, using any fitting method, the Dp and Fp had small bias and wCV repeatability. Conclusion: To accurately measure the IVIM parameter on Unity MR-Linac for the head and neck, it is necessary to consider increasing the number of b-values to reduce the bias of the estimated parameters compared to increasing NSA. Moreover, the unsupervised deep learning approach outperforms LM regression and polynomial Bayesian estimation when measuring IVIM parameters for multi-b-value DWI.
