상세 보기
초록
Gliomas, the most common and aggressive primary brain tumors, exhibit substantial morphological complexity and tissue heterogeneity. Accurate segmentation from Magnetic Resonance Imaging (MRI) is critical for diagnosis, treatment planning, and prognosis. Despite recent advances in automated approaches, challenges persist in jointly achieving high 3D segmentation accuracy, computational efficiency, and interpretability. To address these, this study proposes MSK-Net (Multi-Scale KAN U-Net), a novel 3D segmentation framework that incorporates the Kolmogorov-Arnold Network (KAN). The architecture introduces a KAN-Boosted Attention Module (KBAM) for enhanced feature representation and a Cross-Scale Gating Module (CSGM) to enable efficient multi-scale feature fusion. Comparative evaluation of KAN basis functions demonstrates that Bsplines provide the most favorable trade-off between accuracy and efficiency in delineating complex tumor boundaries. Extensive experiments on BraTS2018, BraTS2021, and BraTS-GLI from the BraTS2024 Challenge validate that MSK-Net consistently achieves state-of-the-art performance, outperforming over 20 existing methods. Furthermore, explainability analysis using Score-CAM highlights MSK-Net's strong sufficiency and plausibility. These findings underscore the framework's ability to balance accuracy, explainability, and efficiency, establishing MSK-Net as a promising tool for clinical brain tumor segmentation. Future large-scale clinical validation could further confirm its utility in supporting radiologists to improve diagnostic precision and patient outcomes. Codes will be made available at: https://github.com/kanglzu/msk_net.
키워드
- 제목
- MSK-Net: Multi-scale spatial KANs enhanced U-shaped network for explainable 3D brain tumor segmentation
- 저자
- Wang, Yutong; Kang, Zhongfeng; Fan, Xiaopeng; Wu, Zijin; Yang, Shantian; Zhao, Qinghua; Song, Zichen
- 발행일
- 2026-07-08
- 유형
- Article
- 권
- 346
- 언어
- ENG
- 출판사
- ELSEVIER
- 발행국가
- 네덜란드
- ISSN
- E 1872-7409
P 0950-7051