In my Thesis, I explored and tried different methods for image segmentation with various image sources from diverse microscope types, in order to be able to identify, localize and segment nuclei at pixel level. My goal was to start with the so-called classical computer vision algorithms and then shift that knowledge and experiments to deep learning solutions with neural networks. I implemented and evaluated methods, so in the future these can be the basis of helper applications for research.
All the discussed implementations had to adapt to different types of images on a limited dataset, where generalization is hard, because of the changing nuclei representations from image to image. However, that way we have the freedom to use only one model for each case, and we do not need to create different solutions.
Apparently, we can solve the problem with the classical methods, but with the use of neural networks we achieved higher accuracy which is important in medical applications. All this led to the conclusion, that the work of my thesis could be used for automatic medical image analyzation.