Attention Seekers U-Net with Mamba for Sub-cellular Segmentation

Abstract

Accurate segmentation of subcellular structures from microscopy images is crucial for understanding cellular processes and functions, but it presents significant challenges due to factors such as noise, low signal-to-noise ratios, limited resolution, and complex spatial arrangements. To address these challenges, we introduce CMU-Net, a novel hybrid architecture that combines the strengths of U-Net, Mamba blocks (SSMs), and Convolutional Block Attention Modules (CBAM). U-Net provides a strong foundation for feature extraction, Mamba blocks efficiently capture long-range dependencies, and CBAM modules refine feature representations by selectively focusing on relevant information. We evaluated CMU-Net on three diverse datasets consisting both fluorescence and label-free microscopy images of mitochondria and endoplasmic reticulum (ER). The quantitative and qualitative results demonstrate that CMU-Net consistently outperforms various baseline methods, including established CNN-based and Transformer-based models, achieving improved segmentation accuracy and boundary representation. This study highlights the potential of our hybrid approach to significantly contribute to the field of subcellular image analysis, promoting a deeper understanding of cellular organization and function. Code is available at https://github.com/beasthunter758/CMU-Net.