Doctoral defence: Mohammed Abdulhameed Shaif Ali "Deep Learning Methods for Cell Microscopy Image Analysis"

On 21 May at 10:15, Mohammed Abdulhameed Shaif Ali will defend his thesis "Deep Learning Methods for Cell Microscopy Image Analysis" to obtain the degree of Doctor of Philosophy (in Computer Science).

Supervisors: 
Lecturer Dmytro Fishman, University of Tartu 
Associate Professor Leopold Parts, University of Tartu 

Opponents: 
Professor Carolina Wählby, Uppsala University and SciLifeLab (Sweden)
Assistent Professor Mohammed Al-masni, Sejong University (South Korea) 

Summary
Analyzing microscopy images of cells is central to modern biological research as it allows scientists to map cellular activity and enables developing new drugs. With the advent of advanced microscopes and the increasing amount of microscopy data, there is a growing demand for automated and accurate image analysis procedures. Deep learning, a type of artificial intelligence, has shown promise in recognizing structures in images and performing tasks like detecting objects within them. In this thesis, we explored and implemented various deep-learning methods to analyze cell microscopy images in a series of studies that comprise the core contributions of this thesis.

We first focused on delineating cells' nuclei (i.e., segmentation), which is the preliminary step in microscopy image analysis. This task is challenging, particularly for brightfield images where nuclei are not marked in a separate acquisition. To do so, we initially evaluated cutting-edge deep-learning methods for nuclei delineation. We then proposed a lightweight model, PP-U-Net, for this purpose. PP-U-Net performs as well as the best available techniques, despite being 20 times smaller in size. During investigating the segmentation errors of the model, we found that the mistaken images often contain artifacts. To tackle this challenge, we next developed a novel framework, ScoreCAM-U-Net, which utilizes deep learning techniques to identify and mitigate artifacts with minimal human input. ScoreCAM-U-Net has the ability to outline objects in fine detail (at pixel level) with the simplicity of using whole image labels as input, making it a potential standard step for all large-scale microscopy imaging experiments. Finally, we used the knowledge gained from this work to support drug candidate screening research. The methods developed in this thesis have been utilized by a variety of industrial and academic organizations, which have benefited from them in their work.