A novel deep neural network-based framework for 3D glioma tumor data segmentation is introduced in this research work. The proposed technique, which integrates clinical data and a physics-inspired regularizer into a deep learning model, accurately identifies glioma sub-regions from the multi-modal MRI scans. This data-informed segmentation approach augments a 3D U-Net model with a lightweight clinical-metadata encoder used to modulate features via feature-wise linear modulation (FiLM), and a learnable non-linear 3D reaction-diffusion PDE that refines class probabilities, with a focus on the Enhancing Tumor (ET) region. It uses routine clinical variables to inform representation learning and regularizes predictions with a compact, edge-aware dynamics prior. The results of the performed simulations and method comparisons are finally discussed here. The high performance metric scores achieved by the described segmentation framework illustrate its effectiveness.