Lightweight Brain Tumor Segmentation Through Wavelet-Guided Iterative Axial Factorization Attention.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate 3D brain tumor segmentation is crucial but challenging for current deep learning models.
• Conventional methods often suffer from high computational costs or ineffective multi-scale feature representation.

Purpose of the Study:

• To develop a lightweight deep learning framework for efficient and accurate 3D brain tumor segmentation.
• To address the limitations of existing methods in terms of computational overhead and multi-scale feature handling.

Main Methods:

• Employed adaptive discrete wavelet decomposition to capture multi-scale information and preserve fine details.
• Utilized iterative axial factorization attention to reduce computational complexity in 3D spatial modeling.
• Developed a lightweight deep learning framework integrating these novel modules.

Main Results:

• Achieved high performance on BraTS2020 and FeTS2022 datasets with Dice scores of 85.0% and 88.1%, respectively.
• The model uses only 5.23 million parameters and 9.75 GFLOPs, demonstrating superior efficiency.
• Effectively balances segmentation accuracy with computational efficiency, outperforming state-of-the-art methods.

Conclusions:

• Integrating frequency-domain analysis with optimized attention mechanisms enhances medical image segmentation.
• The proposed framework offers a scalable and high-performance solution for clinical brain tumor diagnosis.
• Paves the way for broader clinical applications of advanced medical image segmentation algorithms.