SegAnyNeuron: a neural image segmentation network with strong generalization performance by modeling image intensity variation

Neuron reconstruction is a critical step in obtaining quantitative parameters of fine neuronal morphology from microscopic imaging data. Deep neural networks have been extensively applied in this field, with the predominant approach focusing on accurately identifying the voxels occupied by neurites and enhancing the image intensity of these identified voxels to improve the accuracy of three-dimensional reconstruction. However, significant domain shifts in image intensity between foreground and background constrains the generalization performance of existing segmentation models across multi-modal datasets. Consequently, a universal representation framework with cross-domain robustness has yet to be established. In this study, we leverage the directional properties of neurites within neuronal images to extract spatial orientation features at the voxel scale. These features are integrated through multi-channel representation, which constrains the network's spatial perception of neuronal morphology, thereby significantly reducing the impact of image domain discrepancies on model adaptability. Experimental results demonstrate that our proposed method improves F1-Scores by 7.9%-22.8% compared to the baseline UNet-3D model and outperforms existing state-of-the-art approaches by 6.0%-9.1% in F1-Scores when evaluated on cross-modal neuronal image datasets. This work provides a novel biologically informed solution to address domain generalization challenges in neuronal image segmentation.