NeuroMamba: Multi-Perspective Feature Interaction with Visual Mamba for Neuron Segmentation

TL;DR

NeuroMamba tackles neuron segmentation in volumetric EM by unifying patch-free, global voxel-level modeling with precise local boundary cues. It introduces the MPFI block, combining BDFE for fine-grained boundaries, SCFE for resolution-aware global dependencies, and CFI for dynamic cross-modulation fusion, enhanced by a resolution-prior-based scanning strategy. The approach achieves state-of-the-art results across four public EM datasets, including notable ARAND gains on CREMI-A, and demonstrates robustness to both isotropic and anisotropic resolutions. This work offers a scalable, efficient framework for accurate neuron reconstructions and provides code to facilitate adoption in connectomics workflows.

Abstract

Neuron segmentation is the cornerstone of reconstructing comprehensive neuronal connectomes, which is essential for deciphering the functional organization of the brain. The irregular morphology and densely intertwined structures of neurons make this task particularly challenging. Prevailing CNN-based methods often fail to resolve ambiguous boundaries due to the lack of long-range context, whereas Transformer-based methods suffer from boundary imprecision caused by the loss of voxel-level details during patch partitioning. To address these limitations, we propose NeuroMamba, a multi-perspective framework that exploits the linear complexity of Mamba to enable patch-free global modeling and synergizes this with complementary local feature modeling, thereby efficiently capturing long-range dependencies while meticulously preserving fine-grained voxel details. Specifically, we design a channel-gated Boundary Discriminative Feature Extractor (BDFE) to enhance local morphological cues. Complementing this, we introduce the Spatial Continuous Feature Extractor (SCFE), which integrates a resolution-aware scanning mechanism into the Visual Mamba architecture to adaptively model global dependencies across varying data resolutions. Finally, a cross-modulation mechanism synergistically fuses these multi-perspective features. Our method demonstrates state-of-the-art performance across four public EM datasets, validating its exceptional adaptability to both anisotropic and isotropic resolutions. The source code will be made publicly available.

Figures (6)

• Figure 1: Illustration of the standard EM neuron segmentation pipeline and limitations of existing baselines. (a) The standard two-stage pipeline: deep learning-based affinity prediction followed by post-processing. (b) Complementary failures in affinity prediction. The CNN baseline misses ambiguous membranes due to limited context (red box), while the Transformer baseline suffers from boundary imprecision due to the loss of voxel details caused by patch partitioning (green boxes).
• Figure 2: Architecture of the proposed NeuroMamba. (a) The overall pipeline of NeuroMamba. (b) Details of the Encoder Block, where we introduce the MPFI block for multi-perspective modeling. (c) Details of the Decoder Block.
• Figure 3: Illustration of the Multi-Perspective Feature Interaction (MPFI) block. The MPFI block includes two feature extractors and a feature interaction module: (a) Boundary Discriminative Feature Extractor (BDFE), (b) Spatial Continuous Feature Extractor (SCFE), and (c) Cross Feature Interaction (CFI). Our MPFI block models neuronal affinity information from both local boundary discriminative features and global spatial continuity features.
• Figure 4: 2D visualization results on AC3/AC4 and CREMI-A data. The left side shows the EM raw image, and the right side shows the affinity and segmentation results. The red and yellow boxes indicate merge and split errors, respectively
• Figure 5: We evaluate the performance of our model under different hyperparameter configurations on three datasets. The top row reports the VI metric, while the bottom row presents the ARAND metric. The best-performing baseline method and its performance are highlighted in the figure.