• Developing a Multimodal Deep Learning Pipeline for Automated Glioma Subregion Segmentation and 3D Reconstruction with Integrated Spatial Analysis for Clinical Insight

Gliomas, especially grade IV gliomas (glioblastomas), are among the most aggressive brain tumours where their diagnostic and therapeutic precision is often hindered due to the heterogeneous nature of the tumour microenvironment. Current clinical workflows rely heavily on manual segmentation of the whole tumour and its subregions from MRI
scans. This can be a very time-consuming task with high inter- and intra- operator variability. In this report, an end-to-end deep learning pipeline is presented, designed to automate glioma subregion segmentation and 3D tumour reconstruction with morphometric analysis of the subregions. This pipeline aims to improve diagnostic clarity and aid in surgical planning. Through 3 ablation studies, this work evaluates the impact of transfer learning strategies, attention mechanisms and loss optimisation on segmentation performance. The final configuration achieved a Dice score of 0.944 (whole tumour) and 0.917 (enhancing tumour) which demonstrates high segmentation accuracy. Furthermore, the pipeline reconstructs the tumour within the brain in 3D using Marching Cubes algorithm and quantifies its morphometrics. For instance, the distance of tumour centroid to brain midline to aid in clinical decision-making. This work contributes to a scalable and clinically meaningful tool for brain tumour diagnosis as well as more personalised glioma treatment.