Lightweight Brain Tumor Segmentation Through Wavelet-Guided Iterative Axial Factorization Attention
Background/Objectives: The accurate and efficient segmentation of brain tumors from 3D MRI data remains a significant challenge in medical imaging. Conventional deep learning methods, such as convolutional neural networks and transformer-based models, frequently introduce significant computational o...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-06-01
|
| Series: | Brain Sciences |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2076-3425/15/6/613 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background/Objectives: The accurate and efficient segmentation of brain tumors from 3D MRI data remains a significant challenge in medical imaging. Conventional deep learning methods, such as convolutional neural networks and transformer-based models, frequently introduce significant computational overhead or fail to effectively represent multi-scale features. Methods: This paper presents a lightweight deep learning framework that uses adaptive discrete wavelet decomposition and iterative axial attention to improve 3D brain tumor segmentation. The wavelet decomposition module effectively captures multi-scale information by breaking it down into frequency sub-bands, thereby the mitigating detail loss often associated with standard downsampling methods. Ablation studies confirm that this module enhances segmentation accuracy, particularly in preserving the finer structural details of tumor components. Simultaneously, the iterative axial factorization attention reduces the computational burden of 3D spatial modeling by processing attention sequentially along individual axes, preserving long-range interdependence while consuming minimal resources. Results: Our model performs well on the BraTS2020 and FeTS2022 datasets with average Dice scores of 85.0% and 88.1%, with our competitive results using only 5.23 million parameters and 9.75 GFLOPs. In comparison to state-of-the-art methods, it effectively balances accuracy and efficiency, making it suitable for resource-constrained clinical applications. Conclusions: This study underscores the advantages of integrating frequency-domain analysis with optimized attention mechanisms, paving the way for scalable, high-performance medical image segmentation algorithms with broader clinical diagnostic applications. |
|---|---|
| ISSN: | 2076-3425 |