From BraTS Challenges to an Extended Glioma Dataset: State-of-the-Art BrainSegFounder Model Optimization and a Decade of Insights into Multi-Class Glioma Tumor Segmentation.

Recent advancements in artificial intelligence have profoundly impacted medical imaging, par- ticularly in brain tumor segmentation through MRI. This thesis delves into the evolution and challenges of this field over the past twelve years, with a focus on the BrainSegFounder model. This cutting-edge model, based on the SwinUNETR architecture, excels in segmenting pre- treatment glioma tumors into specific classes with remarkable accuracy. Inspired by the pivotal role of the Brain Tumor Segmentation (BraTS) challenges since 2012, which have continu- ally benchmarked and refined segmentation models, our research includes a detailed review of BraTS datasets spanning from 2012 to 2024. We then adapted and validated the BrainSeg- Founder model for the latest BraTS Adult Glioma Dataset, establishing our own baseline through fine-tuning to reproduce and refine results. Our research also investigates various optimization techniques, including the introduction of novel loss functions, adjustments in classification labels, and the implementation of Test-Time Augmentations (TTA) to boost inferential robustness. Despite these enhancements, the model demonstrated significant stability, indicating the challenge in further improving its already high performance. Additionally, we introduced the University Hospital of Padova Dataset for testing and model adaptation. Originally structured as a four-class system, this dataset was modified to align with the standard three-class system employed in the current BraTS challenges, facilitating direct comparisons and model fine-tuning on both skull-stripped and non-skull-stripped versions. The results obtained are promising, indicating the potential for further improvements in the model’s performance. A further adaptation to the original four-class system of the Padova dataset yielded promising yet limited results. Overall, our findings confirm the robustness of the BrainSegFounder model, which, trained on BraTS’s extensive and diverse dataset, showcases remarkable adaptability and potential to enhance segmentation accuracy on external datasets. These findings underscore the significant clinical utility of advanced neural networks in medical imaging, demonstrating their potential for enhancing precision medicine practices. This research contributes profoundly to the broader discourse on precision medicine, affirming the importance of continued innovation in this field.