LATUP-Net: A lightweight 3D attention U-Net with parallel convolutions for brain tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate 3D brain tumor segmentation from MRI is vital for early treatment but challenged by tumor heterogeneity and resource constraints.
• Efficient and accessible medical imaging solutions are needed, especially in developing countries, to meet energy sustainability targets.

Purpose of the Study:

• To introduce LATUP-Net, a Lightweight 3D ATtention U-Net with Parallel convolutions, designed for efficient and high-performance brain tumor segmentation.
• To reduce computational requirements significantly while maintaining precise tumor delineation in MRI scans.

Main Methods:

• Developed LATUP-Net, integrating parallel convolutions for multi-scale feature representation and an attention mechanism for selective feature recalibration.
• Evaluated LATUP-Net on the BraTS 2020 and BraTS 2021 datasets, assessing segmentation performance using Dice and Hausdorff distance metrics.
• Analyzed model interpretability using gradient-weighted class activation mapping and confusion matrices.

Main Results:

• LATUP-Net achieved high average Dice scores (e.g., 90.29% for whole tumor on BraTS 2021) and improved boundary delineation via Hausdorff metrics.
• The model requires significantly fewer parameters (3.07 M) and GFLOPs (15.79) compared to state-of-the-art models, running on a single consumer-grade GPU.
• Interpretability studies showed attention mechanisms aid small region segmentation, highlighting the need to balance local and global features for optimal accuracy.

Conclusions:

• LATUP-Net offers a computationally efficient and accurate solution for 3D brain tumor segmentation from MRI.
• Its low resource demand makes it a promising tool for clinical applications, particularly in resource-limited environments.
• Balancing local and global feature extraction is key for achieving the most precise tumor delineation.