Supervisors: Alvaro Gomariz and Prof. Orcun Goksel
Cell detection methods in microscopy images are key for diagnosis applications as well as advances in biomedical research. In spite of the advances in computer vision, most of these routines still rely on manual analysis. In this project, we explore the use of convolutional neural networks for this task, namely by using 2D and 3D adaptations of U-Net and Siamese networks. While the former focuses on extracting features that are useful for the detection of the labeled cells, the latter learns features that are useful for similarity learning. In this way, template cells can be used to match similarities in a search image. Results in the employed dataset show that 3D models lead to better detection of cells than their 2D counterpart, and that Siamese networks are comparable to the U-Net. A potential caveat with Siamese networks was observed, namely that they can learn fixed embeddings of a template under special circumstances. We explored some solutions to this problem and discuss novel problem settings where Siamese networks can be helpful.