Development of Physics-informed nnU-Net for Glioblastoma Magnetic Resonance Image Segmentation

Abstract

Glioblastoma (GBM) segmentation is particularly challenging due to the highly heterogeneous nature of the tumor. GBM exhibits significant variability in shape, size, and texture across different patients, making it difficult for models to generalize effectively [1, 2]. Moreover, the tumor consists of multiple subregions, including the necrotic core (NCR), enhancing tumor (ET), and edema (ED), each of which appears differently on MRI scans. These subregions do not always have clear boundaries, especially edema, which can infiltrate healthy tissue, making it difficult to distinguish tumor margins precisely. To improve segmentation accuracy, deep learning models like nnU-Net incorporate advanced pre-processing, data augmentation, and post-processing techniques. Standard nn-UNet architectures excel at learning statistical mappings from image intensities to labels, but they can struggle when the test data differ from the training distribution (for example, in different scanners, sequences, motion artifacts, or noise levels). Embedding physics into a nn-UNet could help address these shortcomings. To address this problem, we propose a physics informed nn-UNET model based on the diffusion-reaction equation to guide the tumor segmentation process.