Introduction
Deep learning architectures like U-Net have become the gold standard for medical image segmentation due to their exceptional pattern recognition capabilities. However, these models often lack "biological common sense," treating MRI voxels as purely statistical data while ignoring the physiological laws governing tumor growth and infiltration. This leads to segmentations that may be mathematically optimal but biologically implausible, particularly in data-scarce scenarios or when specific MRI modalities are missing.
This paper introduces a novel biophysics-informed regularization framework that integrates a reaction-diffusion Partial Differential Equation (PDE) into the training process. By forcing models to respect the biological laws of tumor expansion, the proposed method achieves superior accuracy and robustness without adding computational overhead during inference.
Methodology
The core of the proposed approach is a plug-and-play module that can be integrated into any standard segmentation network. The framework, termed Biophysics-Informed Pathological Regularisation, consists of several key components:
The Biophysical Model
The researchers utilize a reaction-diffusion equation to model tumor cell density ($u$). This equation accounts for diffusion (cellular infiltration into neighboring tissue) and reaction (cellular proliferation via a logistic growth model). To bridge the gap between image features and the PDE, a sub-network estimates cell density ($\hat{u}$). Notably, this estimator employs periodic activation functions (Sine), which are significantly more effective than standard ReLUs at capturing the complex spatial derivatives required for physics-informed tasks.
Loss Function and Boundary Conditions
The total loss function combines standard Dice loss with two physics-based terms:
- $\mathcal{L}_{PDE}$: Penalizes violations of the growth equations.
- $\mathcal{L}_{BC}$: Enforces Neumann boundary conditions, ensuring tumor growth respects the physical boundaries of the brain domain.
Crucially, this module is only used during training. Once the weights are regularized, the PDE module is pruned, allowing the final model to maintain standard inference speeds.
Experimental Setup
The method was evaluated using the BraTS 2023 dataset, comprising 1,251 multi-parametric MRI cases (T1, T1c, T2, and FLAIR). Performance was measured across three sub-regions: the enhancing tumor (ET), the tumor core (TC), and the whole tumor (WT). The study tested the regularization across multiple architectures, including U-Net, nn-U-Net, and SegResNet.
Results and Discussion
Quantitative results demonstrate that the biophysics-informed versions consistently outperformed their baseline counterparts across all metrics and architectures.
Accuracy and Precision
As shown in the results table, a standard U-Net's Dice score for the Tumor Core (TC) improved from 90.68% to 91.83% with the addition of biophysical priors. Even more significant gains were observed in boundary precision; the 95th percentile Hausdorff Distance (HD95) for the TC dropped from 6.16mm to 3.99mm, indicating much more precise margin detection.
Resilience to Data Scarcity
A striking finding was the model's robustness. A biophysics-informed model trained on only 30% of the data often outperformed standard models trained on significantly larger datasets. Furthermore, the model maintained high performance even when critical MRI sequences (like T2) were missing, as the biophysical prior helped the network "fill in" missing visual information based on biological growth patterns.
Conclusion
By embedding biological laws directly into the loss function, this framework provides a powerful structural bias that improves segmentation quality and robustness. Future work will focus on incorporating anisotropic diffusion to model how tumors follow white matter tracts and dynamically estimating patient-specific diffusion and proliferation parameters.
Summary updated