Validating the Performance of Brain Tumor Segmentations from Automated Deep Learning Models Using Multiple Raters

This study evaluates the performance of two automated deep learning models for glioblastoma segmentation using MRI data, comparing an independent model trained separately on pre- and post-treatment images with a combined model trained on both. A dataset of 200 unseen cases (100 PreTx, 100 PostTx) was segmented by both models and reviewed by three raters using a structured five-point Likert scale and qualitative feedback. Inter-rater agreement, segmentation similarity (via Dice scores), and failure patterns were analyzed. Results showed both models performed similarly overall, with the independent model yielding more under-segmentation errors, particularly for edema and contrast enhancement. The combined model demonstrated greater robustness and efficiency, suggesting it may be more suitable for integration into research workflows. This work underscores the potential of deep learning models for brain tumor segmentation while identifying key areas for future improvement in model design and evaluation.