Clinical capability of modern brain tumor segmentation models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Deep learning models achieve high performance on the Brain Tumor Segmentation (BraTS) dataset for glioma MRI analysis.
• Concerns exist regarding the generalizability of these models to diverse, real-world clinical MRI data outside the standardized BraTS dataset.
• Previous studies show significant performance degradation in deep learning models when applied across different institutions.

Purpose of the Study:

• To evaluate the cross-institutional applicability and generalizability of state-of-the-art deep learning models on new clinical MRI data.
• To assess the performance of a 3D U-Net model trained on the BraTS dataset when applied to in-house clinical data with varying tumor types and imaging characteristics.
• To compare the model's segmentation accuracy against expert annotations on clinical data.

Main Methods:

• A state-of-the-art 3D U-Net model was trained using the standard BraTS dataset, which includes low- and high-grade gliomas.
• The trained model was evaluated for automatic brain tumor segmentation on an in-house clinical dataset.
• The clinical dataset featured MRIs with diverse tumor types, resolutions, and standardization protocols, differing from the BraTS dataset.

Main Results:

• The model achieved average Dice scores of 0.764 (whole tumor), 0.648 (tumor core), and 0.61 (enhancing tumor) on clinical MRIs.
• These scores surpass previously reported results for cross-institution and same-institution datasets using different methodologies.
• Performance on clinical data, while lower than on BraTS data, demonstrated impressive segmentation capabilities on unseen images from a separate clinical institution, comparable to inter-expert variability.

Conclusions:

• State-of-the-art deep learning models exhibit promising cross-institutional predictive performance for brain tumor segmentation.
• These models significantly outperform previous methods and can effectively transfer knowledge to new brain tumor types without additional model retraining.
• The findings support the clinical utility of deep learning models trained on curated datasets for analyzing diverse, real-world medical imaging data.