Physics-Informed Neural Network for Estimating Cellular Microenvironment Parameters In Radiotherapy Treatment Response Assessment
Abstract
Purpose
Diffusion-weighted MRI (dMRI) is promising for estimating tumor microenvironment characteristics from measured signals that encode the microscopic diffusion of water molecules constrained by cellular architecture. Estimation accuracy is compromised by low signal-to-noise ratios. While deep learning-based end-to-end mapping can improve performance, they suffer from limited interpretability. This study proposes a physics-informed neural network (PINN) framework to enhance cell parameter estimation accuracy while preserving explainability.
Methods
A parameter-to-signal (Para2Sig) model was first trained to represent the biophysical relationship from cell parameters to dMRI signals using data generated by theoretical dMRI model. The PINN framework minimized losses in both dMRI signal domain and cell-parameter domain to establish a signal-to-parameter (Sig2Para) prediction model. The pre-trained Para2Sig was integrated to the PINN framework as a constraint of the dMRI physical model, and in lieu of the theoretical dMRI model to enable error backpropagation using auto-differentiation approach. The entire framework was trained in a self-supervised scheme using data generated on the fly. We evaluated the effectiveness of the proposed framework in simulated data and experimental data with Jurkat cell line. A simulated brain tumor dataset was further employed to qualitatively assess the performance in a realistic clinical scenario.
Results
Pre-trained Para2Sig model reproduced ground-truth dMRI signals with 99.92% accuracy. Incorporating Para2Sig as a physical constraint, the PINN reduced prediction errors to 3.5%, 5.3%, and 8.4% for cell diameter (d), intracellular volume fraction (Vin), and extracellular diffusion coefficient (Dex), respectively, outperforming the baseline model trained without physics constraint (5.1%, 8.0%, 10.6%). On experimental data, PINN yielded a d estimation of 9.19 µm (ground truth: 10 µm), compared with 14.23 µm from the baseline. On the brain tumor dataset, PINN achieved PSNR/SSIM values of 35.51/0.975, 30.91/0.978, and 23.60/0.782 for d, Vin, and Dex, respectively.
Conclusion
The proposed PINN framework improved cell parameters estimation accuracy and interpretability.