This is the open-source code for the paper "NexMamba:Nexus-driven Multi-domain Mamba Network for Medical Image Segmentation." The paper is currently under review.

Abstract : Current Mamba-based serializable long-range dependency modeling models have been widely applied to medical image segmentation. However, such models lack structural constraints during cross-layer feature fusion, readily leading to semantic gaps between shallow and deep features while struggling to capture high-frequency details within images. To address these deficiencies, we propose NexMamba, a U-shaped segmentation framework that deeply integrates Mamba encoders, graph-structured interactions, and frequency-domain detail enhancement. Specifically, we introduce the Cross-Scale Interaction Module (CGF-Nexus), which explicitly models traditional skip connections as sparse topological graphs. This achieves efficient multiscale semantic alignment through a recurrently interwoven interaction mechanism between graph nodes. Subsequently, an adaptive enhancement module(SSEN) is designed for critical high-frequency components such as edges and textures, enabling frequency-domain detail augmentation. Finally, we introduce an improved segmentation head(SFF-Head) that fuses spatial and frequency domains, structurally decoding local high-frequency details alongside global low-frequency semantics to generate finer, more continuous, and anatomically consistent predictions. Extensive experiments across public datasets including Synapse, ACDC, and ISIC2017/2018 demonstrate that results demonstrate that NexMamba achieves Dice coefficients of 86.03% and 92.75% on the Synapse and ACDC datasets respectively, alongside 88.85% and 89.53% on the ISIC2017 and ISIC2018 datasets. This surpasses existing state-of-the-art methods, striking a favourable balance between segmentation accuracy, generalisation capability, and computational efficiency.