Numerical integration in image processing – Combining computer vision with Runge-Kutta methods – Nikomidisz Eftimiu

In modern image processing, end-to-end machine learning pipelines have become some of the most popular tools at our disposal. Feeding raw data to a neural network and receiving a solution sounds fantastic in theory, but this paradigm is not perfect in practice; computational costs, data requirements, and explainability are all hurdles that we’ve yet to overcome.

My PhD. research asks the question whether it would be possible to avoid these issues altogether by combining shallow, lightweight networks with classical methods. Numerical analysis is a field of mathematics that is commonly used to study systems of differential equations, but its applications in image processing extend beyond diffusion or deconvolution. Perhaps the methods used to analyse physical phenomena could also be employed to detect or segment cells. Perhaps they could be leveraged to train smaller networks on less data, if the problem is formulated correctly. My aim is to prove these claims by competing in the Cell Tracking Challenge using hybrid systems that combine popular neural networks with Runge-Kutta methods.

During the seminar, I plan to give some insight into my work in Czechia, and to say a few words about what the PhD. life has been like at Masaryk University for the past 3.5 years.

About Nikomidisz Eftimiu: Nikomidisz is a guest PhD student from Masaryk University working in Carolina Wählby’s lab.