MRI brain tumor segmentation using residual Spatial Pyramid Pooling-powered 3D U-Net.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Neuroscience
• Machine Learning for Healthcare

Background:

• Automated brain tumor segmentation is crucial for efficient diagnosis, reducing lengthy manual processes.
• Traditional U-Net models for semantic segmentation face limitations in contextual information transfer during upsampling.
• Existing advanced methods often increase model complexity and training time.

Purpose of the Study:

• To propose an improved U-Net architecture, SPP-U-Net, for enhanced brain tumor segmentation.
• To increase the contextual scope of segmentation by integrating Spatial Pyramid Pooling (SPP) and Attention mechanisms.
• To achieve comparable or superior segmentation performance without increasing trainable parameters.

Main Methods:

• Replaced standard residual connections in U-Net with a combination of Spatial Pyramid Pooling (SPP) and Attention blocks.
• SPP aggregates multi-scale feature information from different downsampling layers.
• Attention mechanisms integrate local features with global dependencies for improved contextual understanding.

Main Results:

• The proposed SPP-U-Net achieves comparable results to existing complex models.
• The model maintains performance on large dimensions (160 × 192 × 192) without increased trainable parameters.
• Achieved an average Dice score of 0.883 and a Hausdorff distance of 7.84 on Brats 2021 cross-validation.

Conclusions:

• SPP-U-Net effectively enhances brain tumor segmentation by improving contextual information flow.
• The proposed architecture offers an efficient alternative to computationally intensive methods.
• This approach demonstrates significant potential for accelerating brain tumor diagnosis through accurate automated segmentation.