Author: Yan Dai, Jie Deng, Xun Jia, Wen Li 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Johns Hopkins University, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Cell microstructure information is critical for radiotherapy response assessment. Diffusion MRI (dMRI) offers great potential in deriving cell parameters. Yet parameters estimated in existing models via inverse optimization frequently suffer from model robustness issue. This study investigates robustness of the microstructure estimation model and develops a robust approach.
Methods: We considered dMRI scan with pulsed-gradient spin-echo (PGSE) sequence and oscillating-gradient spin-echo (OGSE) sequence with cycle number N=1 and N=2. Using dMRI signal model, we computed signals with cell diameter (d, 6-20μm), intracellular diffusion coefficient (Din, 0.2-3.38μm²/ms), intracellular volume fraction (Vin, 0%-100%), extracellular diffusion coefficient (Dex, 0.2-3.38μm²/ms), and slope of extracellular diffusion coefficient with respect to oscillation frequency (βex, 0-10μm²). We computed Jacobian of signals across the entire parameter space to evaluate sensitivity of signal with respect to microstructure parameters, and derived estimated parameter uncertainty given practical setting of signal-to-noise-ratio (SNR) 20. To build a model to estimate parameters, we applied logarithmic transformation to signals and used principal-component analysis to reduce dimensionality. Three models were constructed (linear regression, polynomial regression, and a 4-layer neural network) to map the dimension-reduced signal to robustly derivable microstructure parameters. We further validated the model in experiments conducted on an 1.5T MR scanner with MC38 cell line.
Results: The d, Vin, and Dex were found robust. Under SNR=20, the uncertainty of estimated d, Vin, and Dex were 3.9μm, 6.4%, and 0.36μm2/ms, respectively. Among the models considered, the 4-layer neural network obtained the best estimation, with MAE of 1.7μm for d, 5.06% for Vin and 0.28μm2/ms for Dex. The experimental study further demonstrated effectiveness of our estimations, with cell diameter difference in ~1μm.
Conclusion: d, Vin, and Dex can be robustly estimated using dMRI due to mathematical properties of the dMRI model. The proposed parameter estimation method holds potential to non-invasively assess microstructure response to radiation.