Kutatószeminárium

Imaging, identifying, and studying cells and their subcellular components are one of the main drivers of research in biology and drug discovery. Image segmentation is an important part of this, as it enables researchers to analyze the number of cells, their shapes, phenotypes or physiological state on microscopic images. For segmentation tasks, convolutional neural networks yield the best results nowadays. In my presentation, I am going to introduce FourierDist, a network, which is a modification to the popular StarDist and SplineDist architectures, which have been proven to work very well on biological images. While StarDist and SplineDist describe an object by the lengths of equiangular rays and control points respectively, our network utilizes Fourier descriptors, predicting a coefficient vector for every pixel on the image. First, I will give you a brief and intuitive introduction to Fourier descriptors, and then go into the details of the architecture of the FourierDist network. I will demonstrate the differences between StarDist, SplineDist and FourierDist on real and synthetic images, and show that Fourier descriptors can achieve higher accuracy than SplineDist and yield the same segmentation quality as StarDist for a high number of parameters, while also outperforming it for a low number of parameters. If the objects to be segmented are not star-shaped, StarDist performs suboptimally compared to SplineDist and the proposed model as well.