A multiscale spectral embedding network with dual–domain attention for cell segmentation and detection

Segmenting objects accurately and reliably is a challenging task for object detection. Tiny objects, like cells, their count, shape, and distribution reveal important information for assisting disease diagnosis in medical and agricultural practices. In this research, we propose a novel Multiscale Spatio-Frequency channel attention (MSFCA) incorporated into UNet for cell segmentation. MSFCA-UNet utilizes frequency and intensity information to improve the network’s capacity for cell identification. The MSFCA module integrates frequency attributes using a Spatial Frequency Network (SFN), followed by a Spatial Channel Attention Network. SFN emphasizes semantic information by focusing on high-frequency features that capture important details, such as the texture and edges of the cell, while the spatial channel attention network recalibrates feature maps by emphasizing the important spatial and channel information. The MSFCA module suppresses irrelevant information and preserves gradient and semantic information. The feature maps at different encoder stages are further fused at the decoder to capture cell boundaries for pixel classification. Experimental results demonstrate that the MSFCA-UNet outperforms traditional convolutional neural networks (CNNs), achieving 99.08% accuracy in cell segmentation and a 0.952 score of the dice coefficient. The findings highlight the efficacy and reliability of our methodology in accurately analyzing hive cells.