Does Anatomical Contextual Information Improve 3D U-Net-Based Brain Tumor Segmentation?

Area of Science:

• Medical image analysis
• Artificial intelligence in neuro-oncology

Background:

• Automatic brain tumor segmentation is crucial for treatment planning.
• Convolutional neural networks (CNNs) show promise in segmenting tumor regions from MR images.
• Context-aware AI is an emerging approach to enhance deep learning for medical imaging.

Purpose of the Study:

• To investigate if incorporating contextual brain anatomy information improves U-Net-based brain tumor segmentation.
• To evaluate the impact of anatomical masks (white matter, gray matter, cerebrospinal fluid) on segmentation accuracy.
• To assess performance in terms of accuracy, training time, domain generalization, and handling of limited MR modalities.

Main Methods:

• Two standard 3D U-Net (nnU-Net) models were trained using MR images plus anatomical contextual information (binary masks or probability maps) as extra channels.
• A baseline model used only conventional MR image modalities.
• The BraTS2020 dataset was used for training and testing on 125 subjects.

Main Results:

• No statistically significant difference in Dice scores was observed between the baseline and contextual information models (p > 0.05).
• Contextual information did not improve model training time or domain generalization.
• Significant improvement (p < 0.05) in whole tumor segmentation was noted only when compensating for fewer available MR modalities.

Conclusions:

• Adding anatomical contextual information (WM, GM, CSF masks or probability maps) does not generally enhance U-Net-based brain tumor segmentation performance.
• Contextual information may offer benefits in specific scenarios, such as reducing false positives in low-grade tumors or improving segmentation with limited MR data.
• Further research may explore alternative methods for integrating contextual information to maximize its benefits in medical image analysis.