LiteFreqMamba:lightweight frequency-enhanced Mamba for efficient 3D brain tumor segmentation

Abstract

Introduction State Space Models (SSMs) have demonstrated strong potential for 3D brain tumor segmentation owing to their linear computational complexity. However, conventional Mamba-based models are often limited by spectral bias, which favors low-frequency information while overlooking high-frequency boundary details, as well as by spatial disruption introduced by 1D scanning. Methods We propose LiteFreqMamba, a novel frequency-enhanced architecture for accurate and efficient 3D brain tumor segmentation. LiteFreqMamba is designed to improve boundary representation and spatial modeling through several key components. First, a Decomposed Frequency-Spatial Convolution (DFS-Conv) encoder is introduced to explicitly decompose features and capture high-frequency boundary information in shallow layers, thereby alleviating boundary ambiguity. Second, a Mamba-Attention Hybrid Bottleneck (MAHB) is developed to preserve spatial structure by combining the 2D-Selective-Scan (SS2D) mechanism for linear-complexity spatial mixing with dense self-attention for finegrained pixel-wise dependency modeling. In addition, Frequency-Calibrated Skip Connections (FCSC) are proposed to dynamically suppress noise in high-frequency feature injection, and a Lightweight 3D Convolutional (LWT-3D Conv) decoder is employed for efficient feature reconstruction. Results Extensive experiments on the BraTS2020 and BraTS2021 benchmarks show that LiteFreqMamba surpasses existing efficient segmentation models and achieves a better balance between inference speed and segmentation accuracy. Discussion LiteFreqMamba is designed to improve high-frequency boundary representation and preserve spatial dependencies in Mamba-based architectures. The proposed framework provides an efficient and accurate solution for 3D brain tumor image segmentation.