Multi-Phase and Hierarchical Unsupervised Learning Framework for Glioblastoma Sub-Region Segmentation in MRI Sequences

Glioblastoma (GBM) is the most prevalent and aggressive form of brain cancer, typically associated with a poor prognosis and a median survival time of 15 months. Effective treatment planning and monitoring require precise segmentation of GBM sub-regions in MRI scans, a task traditionally reliant on time-consuming and expertise-demanding manual annotations. Current unsupervised learning methods for GBM segmentation are limited in accurately segmenting tumour sub-regions due to the high variations in tumour morphology and pathophysiology caused by strong heterogeneity. To address these limitations, we propose a novel multi-phase and hierarchical unsupervised learning framework tailored for GBM sub-region segmentation using multiple MRI sequences. Our approach innovates by leveraging intrinsic image features and spatial relationships encoded in MRI data without relying on annotated datasets. Key contributions include a phased training approach for progressive segmentation refinement and a context-based hierarchical loss function to ensure spatial consistency. Our method was evaluated on the BraTS21 dataset and demonstrates superior performance compared to common clustering methods, achieving balanced segmentation across GBM sub-regions. This framework reduces dependency on extensive labelled datasets, paving the way for more efficient and scalable GBM segmentation. Therefore, our framework shows great potential in GBM sub-region segmentation.