Author: Yan Dai, Jie Deng, Xun Jia, Wen Li, Junzhong Xu 👨🔬
Affiliation: Johns Hopkins University, Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center, Department of Radiology, Vanderbilt University Medical Center, Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Cell size is a vital parameter in evaluating the tumor microenvironment, including cell apoptosis and radiotherapy(RT)-induced immune cell infiltration. The IMPULSED(Imaging Microstructural Parameters Using Limited Spectrally Edited Diffusion) method utilizes pulsed/oscillated gradient spin-echo(PGSE/OGSE) diffusion-weighted MRI to extract microstructural parameters, such as cell size and intracellular volume fraction. While it has demonstrated promising results on high-field(≥3.0T) clinical or preclinical scanners, its translation to a clinical 1.5T MRI simulator for RT planning is challenging due to the low signal-to-noise ratio(SNR), constrained scan time for signal averaging, and limited gradient strength. This study applied Bayesian Experimental Design(BED) to optimize the scanning protocol, maximizing information acquisition within clinically feasible scan times.
Methods: A BED framework was implemented to optimize the selection of PGSE and OGSE b-values within scanner constraints, ensuring an optimal acquisition setting under a fixed ~5-minute scan time. The utility of each acquisition setting was defined as the expected reduction in Shannon entropy from the prior to the posterior distribution of microstructural parameters derived from the IMPULSED model. The optimal protocol were identified based on utility maximization. To validate that the results chosen with higher utility values contained the most information, fully connected neural networks were trained on simulated data from various acquisition settings respectively to predict IMPULSED parameters, with accuracy assessed using normalized root mean square error(nRMSE).
Results: Higher utility values corresponded to improved IMPULSED parameter predictions. The optimal combination of b-values (averages) was determined as PGSE:
b = 500(1), 1000(2)
b=500(1),1000(2) s/mm²; OGSEn1:
b = 1000(2)
b=1000(2) s/mm²; OGSEn2:
b = 346(1)
b=346(1) s/mm², outperforming a evenly-distributed b-value setting. A correlation between utility values and network prediction accuracy confirmed effective protocol optimization.
Conclusion: Using BED, we optimized the IMPULSED acquisition protocol for a clinical 1.5T MRI scanner. The optimized protocol enhances parameter estimation accuracy within the constraint of clinically acceptable scan times.