MBMSA-UNet: A Multi-Scale Attention-Based Instance Segmentation Model for Moso Bamboo Cells

Instance segmentation of moso bamboo cells is a critical step in quantitative structural analysis of bamboo materials and plant phenomics research. Moso bamboo tissues are mainly composed of vascular bundles and parenchyma cells. Within vascular bundles, fiber cells exhibit thick cell walls and extremely dense arrangements, whereas vessel cells are characterized by large diameters and complex internal structures. These features frequently lead to blurred boundaries, structural complexity, and local overexposure in microscopic images, making it difficult for traditional segmentation algorithms to achieve stable and accurate results. Although the U-Net has demonstrated outstanding performance in biological microscopic image analysis, its feature extraction capability and boundary recognition stability remain insufficient when dealing with the composite structure of moso bamboo. To address these challenges, this study proposes an improved model based on a multi-scale attention mechanism, termed MBMSA-UNet (Moso Bamboo Multi-Scale Attention U-Net). Building upon the encoder-decoder architecture of U-Net, the proposed model introduces a multi-scale channel-spatial attention block, aiming to handle the pronounced morphological and scale differences among vessels, fibers, and parenchyma cells. By adaptively reweighting features at different scales, the model enhances cross-layer feature fusion and strengthens responses to key regions, thereby effectively suppressing local overexposure interference and emphasizing boundary features between different cell types. Experimental results demonstrate that, compared with the U-Net and several of its improved variants, MBMSA-UNet achieves higher segmentation accuracy and greater robustness on microscopic images of moso bamboo, providing a solid foundation for fine-grained quantitative analysis of complex bamboo tissues.