Multi-scale non-local means with shape prior for enhancement of cell membrane images

Light sheet microscopy allows rapid optical sectioning of whole chicken embryos during early development, visualizing cells through transgenic expression of a fluorescent cell membrane label. Low signal-to-noise ratios and large variability in intensity make segmentation of these images difficult, often resulting in discontinuous cell edges. We propose a multi-scale non-local means with shape prior (MS-NLM-SP) method for image enhancement, which takes into account cell-specific characteristics, such as size, geometry, and periodicity. Investigating theoretical connections between our approach and anisotropic diffusion reveals that the first step, detail-enhanced multi-scale representation, corresponds to selective backward diffusion, while the second step of MS-NLM-SP resembles anisotropic forward diffusion. Experimental results show that incorporating cell-specific features into non-local image denoising methods improves subsequent segmentation.