Deep Learning Approaches for BraTS-Style Brain Tumor Segmentation and External Evaluation on Local Hospital MRI Data

Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) supports diagnosis, treatment planning, and follow-up assessment. Manual segmentation remains time-consuming and dependent on expert interpretation, which has motivated the use of automatic deep learning methods. However, strong performance on standardized benchmark datasets does not necessarily indicate equivalent performance on independent clinical MRI data. This thesis evaluates a MONAI SegResNet-based framework for automatic high-grade glioma segmentation using BraTS-style multimodal MRI. To investigate the effect of training-set size, three SegResNet models were trained using 100, 200, and 403 BraTS cases. Performance was assessed on an unseen BraTS benchmark cohort using Dice, HD95, sensitivity, specificity, and precision. External generalization was then evaluated on an independent local hospital high-grade glioma MRI dataset with expert reference annotations. A mixed BraTS–local training experiment was also performed using 100 BraTS cases and 21 local cases, followed by evaluation on an independent 10-case local test set. The models achieved high performance on the BraTS benchmark, and segmentation accuracy increased with larger BraTS training sets, although the gain became smaller as additional standardized cases were added. In contrast, evaluation on local hospital MRI resulted in lower performance than on BraTS, indicating reduced generalization to independent clinical data. The mixed-training experiment led to only a modest improvement on the independent local test set. These findings indicate that increasing benchmark training data improves segmentation performance within the BraTS domain, while external validation remains necessary to assess performance on clinical MRI. Incorporating local data provided limited benefit in this study, suggesting that larger and more diverse clinical datasets are likely needed to improve model robustness for real-world applications.