Robustness of brain tumor segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Neuro-oncology

Background:

• Brain tumor segmentation is crucial for diagnosis and treatment planning.
• Current deep learning models achieve high performance on benchmark datasets.
• Clinical applicability of these models remains unevaluated.

Purpose of the Study:

• To investigate the generalization performance of deep neural networks (DNNs) for brain tumor segmentation in clinical practice.
• To evaluate state-of-the-art segmentation methods and propose modifications for improved generalization.

Main Methods:

• Evaluation of three state-of-the-art methods, a U-Net architecture, and a cascadic Mumford-Shah approach.
• Implementation of two topology-preserving modifications to enhance generalization.
• Comparative analysis of model performance on benchmark versus clinical data.

Main Results:

• A well-trained U-Net architecture demonstrated the best generalization performance.
• Advanced model extensions were found to be potentially detrimental in realistic scenarios.
• Generalization capabilities of DNNs in medical image analysis are significantly limited.

Conclusions:

• Deep neural networks exhibit severely limited generalization in brain tumor segmentation for clinical practice.
• Existing network topologies are optimized for benchmark datasets, not real-world application.
• Simpler, well-trained models like U-Net may be more suitable for clinical deployment.