Speaker
Description
Mechanical forces shape cells and also influences gene expressions. Precise morphological measurement gives us the insight on the mechanical forces experienced by the cells. Current state-of-art deep-learning segmentation methods, such as Cellpose, gives reliable segmented images, whose boundary is at the pixel resolution. The pixel-size resolution can give rise to more than 10% errors in perimeter and curvature measurement when the boundary is wavy, not straight. We develop an image analysis program, FineBoundary is a GPU-accelerated refinement algorithm, which takes into account the intensity profile of the boundary and achieves subpixel precision in determining cell boundary, improving the accuracy to less than 3% error. The algorithm of the FineBoundary involves steps by refining contour by computing boundary normals, sampling fluorescence intensity profiles on the GPU (CuPy), and fitting Gaussian models in parallel using pygpufit. This iterative process corrects local boundary offsets while preserving overall topology. Vertex detection enables downstream morphometric measurements including perimeter, area, tortuosity, shape index and vertex count. This framework bridges deep-learning segmentation with physics-informed refinement, enabling quantitative studies of cellular mechanics and morphology. We applied this algorithm to study the apical constriction of Madin-Darby Kidney Canine cells and detecting lobes puzzle cells of Arabidopsis.
