HSBI Researcher Develops AI for More Efficient Cell Segmentation

Bielefeld (hsbi). The two images on the computer show the same biological cells. On the left, the microscopic image shows some pancreatic stem cells on a porous substrate. The small white flecks – the individual cells – are different in size and their elongated forms are a little blurry against the grey background. It is not easy to distinguish whether a pixel belongs to the cell or to the background. Eiram Mahera Sheikh directs the visitors’ attention to the right-hand side: the same arrangement of cells, but this time they are in different colours and show clear boundaries to the background. “With this mask, the image is of much better use for analysis. Here, the cells are segmented clearly thanks to the use of AI.” Sheikh wants to make this process even more efficient in the future by using AI to make AI smarter. In other words, she wants to speed up the necessary training of cell segmentation AI by using an AI method.

Valuable for cancer therapy: cell segmentation provides insights into cellular changes

But let’s start at the beginning. After several years in her profession, the 33-year-old computer scientist specialised in Data Science and AI as part of her master’s degree studies. Originally from India, she has now come to Hochschule Bielefeld – University of Applied Sciences and Arts (HSBI) as a doctoral candidate. With cell segmentation, she has chosen a topic that is of great importance for biomedical research and diagnostics: “If we analyse the images of segmented cells, it provides us with important insights into the cells’ characteristics, into pathological changes or into how cells react to certain substances,” Eiram Mahera Sheikh explains. “This information is particularly important in cancer therapy or pharmacological research.” Sheikh’s research, however, starts two steps earlier – on a meta level, so to speak. “Cells can now be segmented by AI. What I am interested in, is training this AI.” And that is a tricky thing indeed ...

Doctorate on cell segmentation is realised as part of the East Westphalian research network SAIL

Professor Dr. Wolfram Schenck is sitting in a meeting room at HSBI. He is listening intently to Sheikh’s presentation and looking at the images projected to the wall as they discuss the doctoral candidate’s interim results. The project is part of the interdisciplinary research network SAIL. This acronym stands for SustAInable Life-cycle of Intelligent Socio-Technical Systems. The network, which is funded by the Land North Rhine-Westphalia with up to 14.8 million euros, is a collaboration of Bielefeld University as the consortium leader, HSBI, OWL University of Applied Sciences and Arts and Paderborn University. Their common goal is to make AI systems work transparently, safely, sustainably and solidly throughout their entire life cycle. For this purpose, they take a 360-degree approach to AI: In addition to conducting fundamental research on AI, the network’s approximately 90 scientists examine the technology’s impact on humans and society and specific applications in industry 4.0 and intelligent healthcare. Wolfram Schenck, Professor of Engineering Computer Sciences, Scientific Director of the Center for Applied Data Science’s (CfADS) Data Analytics Cluster on Gütersloh Campus and HSBI’s spokesperson in the SAIL project, supervises Eiram Mahera Sheikh’s doctoral thesis, which is at the interface of fundamental research and healthcare and aims to provide insights beyond that.

Image data labelling: How can highly specialised experts save time?

Professor Schenck explains the challenges of Sheikh’s work: “In order to learn cell segmentation, the cell segmentation AI needs a lot of labelled data as training material – pixel-level image data that are enhanced with additional information and categorised.” This is an elaborate procedure: for every pixel of an image of biological cells, she must determine whether it belongs to the categories “background,” “membrane” or “cell body.” Professor Schenck goes on to say that this so-called labelling can only be done manually be experts who know exactly what an image shows. Is it a cell membrane, a dead cell or just background? Pixel by pixel, this must be determined by someone who knows what they are doing to make the AI fit for the segmentation of certain cells.

“As this is a time-consuming and monotonous task, experts are usually not too keen on it. In addition, the labelling process is also very expensive, as it consumes so much of these highly trained specialists’ time,” adds Dr. Constanze Schwan, who has joined the meeting. As part of HSBI’s Career@BI programme, the computer scientist teaches at the Faculty of Engineering and Mathematics and works in the research department of the biotech company Miltenyi Biotec in Göttingen in parallel. She thus contributes her knowledge of industry requirements: “It would be a big improvement if we had a model or algorithm that would reduce the efforts we invest in labelling.” Existing segmentation procedures that are suitable for applications such as autonomous driving reach their limits when it comes to microscopic images. One of the reasons for this is that these images have more complex, subtle structures, with the images also including image stacks on a time axis for 3D structures. “These images or image stacks often contain complex structures that are difficult to segment,” Dr. Alaa Othman, the last member of the group, explains. As a junior research group leader in the SAIL network, the electrical engineer supervises Sheikh’s doctorate together with Professor Schenck.

Eiram Mahera Sheikh’s AI is designed to identify image regions that are particularly important for it to learn cell segmentation. This means that image pixels no longer have to be labelled manually in their entirety – only those that are particularly valuable as training data.

A big step forward: Sheikh’s AI sorts out the unimportant training data

Eiram Mahera Sheikh, in turn, relies on AI to avoid the problem of complex structures and to reduce the workload of specialist teams who need to train the cell segmentation AI. “What I intend to do is to identify image regions that are particularly important for the AI to learn cell segmentation. This would mean that image pixels no longer have to be labelled manually in their entirety – only those that are particularly valuable as training data.” And as such a targeted selection of potential training data is called Active Learning and the AI algorithms for cell segmentation come from Deep Learning, the team calls the procedure Deep Active Learning, as Professor Schenck explains.

Sheikh has already identified the training data that are actually valuable: “It’s the most uncertain data – those that the AI algorithm has difficulties with, e.g., due to poor image contrast or due to overlapping or unevenly shaped cells.” In future, only these should be examined and labelled by experts. In clear cases – those in which Sheikh’s AI has been able to label the images –, this effort could become dispensable. Alaa Othman’s summary: “In essence, the cell segmentation AI will need smaller amounts of data and still remain so effective that it will still be able to precisely segment cells.”

Findings can be transferred to other applications such as autonomous driving

The question remains: how can the importance of potential training data be determined? Should the entire image be examined or do local and partial decisions suffice? Soon the group is deep in discussion and non-experts can no longer follow. “Ultimately, the project provides application-oriented fundamental research,” says Professor Schenck. The algorithms developed by Eiram Mahera Sheikh for Deep Active Learning, which divide the image data into regions that are important and unimportant for AI training and thus make cell segmentation more effective and efficient overall, can also be used for other applications such as medical imaging, autonomous driving or the analysis of satellite images. Professor Schenck: “And thus, the SAIL project’s holistic and sustainable approach becomes clear once again.” (uh)