The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS).

Area of Science:

• Medical imaging
• Artificial intelligence
• Neuro-oncology

Background:

• Brain tumor segmentation is crucial for diagnosis and treatment planning.
• Accurate segmentation of gliomas from multimodal MRI is challenging due to tumor heterogeneity and inter-observer variability.
• The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) was established to facilitate objective evaluation of segmentation algorithms.

Purpose of the Study:

• To benchmark the performance of state-of-the-art algorithms for brain tumor segmentation.
• To assess the performance of 20 automated segmentation algorithms against manual annotations by multiple raters.
• To provide a publicly available dataset and evaluation platform for ongoing research in brain tumor segmentation.

Main Methods:

• Twenty tumor segmentation algorithms were applied to 65 multi-contrast MRI scans of glioma patients and 65 simulated scans.
• Manual annotations were provided by up to four raters for ground truth.
• Performance was quantitatively evaluated using metrics such as Dice scores, comparing algorithms to each other and to inter-rater variability.

Main Results:

• Significant inter-rater variability was observed in manual segmentation (Dice scores 74%-85%), highlighting the task's inherent difficulty.
• Individual algorithms showed varying strengths across different tumor sub-regions, with performance comparable to human raters.
• No single algorithm consistently ranked highest across all sub-regions.
• Fusing multiple algorithms via hierarchical majority vote improved segmentation performance above individual methods.

Conclusions:

• Brain tumor segmentation remains a challenging task, even for state-of-the-art algorithms.
• Algorithm fusion offers a promising approach to enhance segmentation accuracy and robustness.
• The BRATS benchmark provides a valuable, ongoing resource for advancing brain tumor image analysis research.