Skip to main content

LUMI provides new opportunities for artificial intelligence research

The supercomputer LUMI is scheduled to begin its operation at CSC’s data center in Kajaani, Finland, in early 2021. When installed, it will be one of the world’s most efficient supercomputers used for research purposes. The huge computational power of LUMI is mainly based on a large number of graphics processing units, or GPU processors. They are particularly well suited for different artificial intelligence (AI) applications, deep learning in particular.

Researcher Kimmo Kartasalo from Tampere University is very familiar with the requirements of AI applications. He is a member of the bioimage informatics group at Tampere University, led by Assistant Professor Pekka Ruusuvuori. In its latest project, the group trained AI to diagnose and grade prostate cancer in cooperation with the Karolinska Institute in Stockholm. The research results showed that the AI system was able to distinguish between cancerous biopsies and benign biopsies almost without error. The study was published in the prestigious The Lancet Oncology journal.

researcher Kimmo Kartasalo
Researcher Kimmo Kartasalo is a member of the bioimage informatics group at Tampere University in Finland. The research group trained AI to diagnose and grade prostate cancer in cooperation with the Karolinska Institute in Stockholm. Image: CSC

Artificial intelligence to support pathologists in their work

In the project, the role of Kartasalo, who is finalizing his doctoral dissertation, was to engage in image analytical development: to develop and test algorithms.

– In the study, digital pathology scanners were used to scan biopsies on microscope slides into huge digital images, and the artificial intelligence analyzed the image data. Thousands of diagnostic biopsies were used as research data, says Kartasalo.

The present practice in health care is that the pathologist processes the slides and, based on his or her assessment, a decision is made on the necessary treatment. However, there is a worldwide shortage of pathologists, and, when it comes to prostate cancer, the need for diagnostics will keep on increasing because of the aging population. Artificial intelligence can be used for creating a tool for pathologists that can improve the work efficiency, while also promoting patient safety by acting as a safety mechanism.

– The research study used an artificial intelligence system based on deep neural networks. To train a neural network, the training algorithm must go through thousands of digital images pixel by pixel. During the training process, the neural network model is adapted to make it as accurate as possible. The accuracy is evaluated against the assessments of experienced pathologists, Kartasalo explains.

Graphics processing units accelerate the computation process

In computational terms, it is a very heavy process to run through thousands and thousands of samples pixel by pixel. In this work, graphics processing units, or GPUs, play an important role.

– The training algorithms that go through images during the training stage of the neural network have been optimized to take advantage of the high parallel computing capacity of GPUs. In practice, when using GPUs, the computing time may drop to one tenth or even less from the time conventional CPU processors would need for the same task.

For its computing tasks, the research group used CSC’s supercomputers: the use of CSC’s earlier Taito cluster reduced the computing time to about two weeks. The final training of the model was done using CSC’s supercomputer Puhti, whose pilot users the research group was. Puhti features even more GPU processors with higher processing speed than the earlier system, which dropped the total computing time to 2–3 days. Using a conventional CPU, it would have taken two to three years to train the artificial intelligence model.

– Everyone can surely understand the difference between waiting for research results for three days instead of three years. It would be impossible to make research if prototyping could not be completed within a reasonable timeline. The actual computational needs were many times higher. The final training phase of the artificial intelligence model was only the tip of the iceberg in the research project that lasted several years – the time was needed for testing different algorithms and design solutions. Using conventional CPUs, this project would probably have been left for the future generations to finalize.

In addition to graphics processors, high-speed hard disk capacity was found to play a key role in the computing, particularly when processing big data.

– If the hard disk is so slow that it cannot keep up with the computing speed of a modern GPU, the overall process is typically as slow as its weakest link, which is typically the hard disk. Therefore, we also need very high-speed hard disks, and Puhti and LUMI meet also this requirement, says Kartasalo.

More complex problems and new research questions

Kartasalo believes that the supercomputer LUMI and its huge computing capacity will open up a lot of different research opportunities, for example, in research utilizing AI.

– If Puhti was a big step compared to the previous level, then LUMI is a giant leap forward. Compared to earlier studies, we can do similar things, only faster. For example, if we can reduce the computing time from three days to three hours, it means that we can test new ideas in an even faster cycle. If we get a new idea we want to test in the morning, some test results will already be ready after lunch. This will accelerate the creative process.

– LUMI also enables posing new kinds of research questions. One trend is to increasingly study these matters from a statistical point of view. Most of the neural network algorithms are stochastic, i.e., the results vary from one run to another. If there is more computing capacity available allowing us to make more runs, it is possible to tackle the uncertainties of the assessments produced by artificial intelligence. For example, for a physician it would be interesting to know not only the diagnosis produced by an AI but also how confident the AI is about its decision. We can do these kinds of things as we get more computing capacity, Kartasalo continues.

Combining data from different sources enables the modeling of more complex problems than before.

– We have also been working on tissue samples related to breast cancer, for example. There are complex molecular-level phenomena behind each image. Artificial intelligence can be used for combining molecular data with the information obtained from the image, and the conclusions can be drawn from this combined data.

– In addition, I also hope that the increased computing capacity could support the introduction of artificial intelligence outside the actual AI research. For us, who are developing AI applications, AI is a tool, but it is also an interesting research subject in itself. Hopefully, the time would be right for applying artificial intelligence in other fields of science as well, let’s say in social sciences, for example, Kartasalo continues.

Finally, Kartasalo encourages other research scientists to already start expanding their thinking and formulating research questions with a view to the supercomputer LUMI.

– Now is a good time to turn ideas into research plans. It would be great if research could be carried out across disciplinary boundaries: if we could begin asking the kind of research questions that we would have liked to have answered earlier but that we haven’t been able to answer. Soon, we will have a fairly unique opportunity to take advantage of this supercomputer and to answer all sorts of questions that have previously seemed impossible. So, let’s make the most of it!


Artificial intelligence for diagnosis and grading of prostate cancer in biopsies: a population-based, diagnostic study

Peter Ström, Kimmo Kartasalo, Henrik Olsson, Leslie Solorzano, Brett Delahunt, Daniel M Berney, David G Bostwick, Andrew J. Evans, David J Grignon, Peter A Humphrey, Kenneth A Iczkowski, James G Kench, Glen Kristiansen, Theodorus H van der Kwast, Katia RM Leite, Jesse K McKenney, Jon Oxley, Chin-Chen Pan, Hemamali Samaratunga, John R Srigley, Hiroyuki Takahashi, Toyonori Tsuzuki, Murali Varma, Ming Zhou, Johan Lindberg, Cecilia Lindskog, Pekka Ruusuvuori, Carolina Wählby, Henrik Grönberg, Mattias Rantalainen, Lars Egevad, Martin Eklund.

The Lancet Oncology, Jan. 8, 2020

Have a look at the video below to see what Kartasalo tells about the possibilities LUMI will bring for research using AI:

This interview is part of an interview series, where different LUMI use cases will be introduced.

Author: Anni Jakobsson, CSC