Whole-body Cell Segmentation of 3D C.elegans Based on Displacement Field and Multi-frequency Feature Fusion

High-precision cell segmentation is an important basis for quantitative analysis of gene expression in C.elegans. In response to challenges such as dense cell distribution and blurred boundaries in microscopic images, this study constructed a deep learning model based on displacement vector field. Multiscale feature fusion was achieved through multi-branch dilated convolution to improve the segmentation accuracy between dense cells. Wavelet convolution was used to construct multi-frequency feature representation, combined with reverse attention to suppress background noise, and the Laplace operator was introduced to improve the distinction of cell boundaries.