Decouple, Collaborate, Match: Prototype-Driven Cognitive Mutual Learning for Brain Tumor Segmentation

Multimodal magnetic resonance imaging (MRI)-based brain tumor segmentation faces two critical challenges: (i) insufficient exploitation of inherent frequency-domain characteristics of MRI signals, which leads to the loss of boundary information, and (ii) the difficulty of balancing local morphological adaptability with global semantic consistency in heterogeneous tumors. To address these issues, we propose a collaborative cognitive segmentation network (CCSNet), a mutual learning framework that enhances the performance of two independent networks: the wavelet fusion network (WFNet) and the adaptive neural attention network (ANANet). WFNet operates in the frequency domain, reconstructing tumor boundaries via multilevel wavelet transforms and tumor-centered radial sampling to explicitly capture high-frequency components. In contrast, ANANet operates in the spatial domain, achieving precise semantic modeling of complex tumor morphologies through adaptive region attention and axial kernel adjustments. To further enhance performance, we introduce a mutual learning mechanism comprising two key components: category-aware prototype matching for addressing intraclass heterogeneity and boundary-aware mutual supervision for enforcing consistency constraints via uncertainty-guided region selection. This mechanism enables bidirectional knowledge transfer between WFNet-ML and ANANet-ML, which are derived from WFNet and ANANet, respectively. Extensive experiments on BraTS 2020 and BraTS 2021 datasets demonstrate that the proposed mutual learning framework consistently improves the performance of both networks. Compared with current state-of-the-art methods, CCSNet reduces the 95th percentile Hausdorff Distance metric by 3.00% and 5.14%, respectively, significantly improving tumor boundary localization.