About a year ago, the LUMI supercomputer – a EuroHPC system based at a CSC datacenter in Kajaani, Finland – debuted in third place on the Top500 list (a position it has maintained on the two subsequent lists). Around that time, CSC announced some of the first workloads that would be hosted on the massive supercomputer. We spoke with Pekka Ruusuvuori, head of the bioinformatics group at the University of Turku and PI of the ComPatAI project, about his team’s research on LUMI.
The research team has been using CSC’s supercomputers for some years: they were a pilot project on Puhti, LUMI’s predecessor, and even used Puhti’s predecessor – Taito – before that. With each new system, the training time – and the scale of training possible – has been whittled down.
But that’s getting ahead of things: first, what is ComPatAI? The name is short for “computational pathology AI,” which is the project in a nutshell.
“Basically, what we are doing with LUMI is to crunch the numbers with the digital pathology data,” Ruusuvuori told HPCwire, adding that “obviously, these days, it’s the artificial intelligence, deep learning-based approach that really enables” high-accuracy analysis. “We can reach a higher accuracy in almost any task where we have enough data compared to anything that has been available before,” Ruusuvuori said.
“What we are trying to do [is] we are trying to build this … general-purpose neural network for extracting features and understanding histological data,” Ruusuvuori explained. (Histology, for reference, is the study of biological tissues.)
In developing this general-purpose computational pathology neural network, Ruusuvuori said there were two main branches of their research. “One is to try to [automate] something that the pathologists are doing routinely at the clinic – so basically to determine whether the patient has a cancer or not,” he said. “And then, if he or she has a cancer, then to determine what is the aggressiveness; what is the grade of the cancer?” Ruusuvuori said that the hope was to provide resources to help experts “work less subjectively, more reliably and also faster with the help of an AI-based decision support tool.”
The other branch: trying to understand disease mechanisms on a deeper level, which – unlike the diagnosis – is not merely streamlining an existing process. Ruusuvuori said that the histopathological images show more information “than what the human eye can really comprehend,” explaining that in his team’s research, they are working to predict gene expression levels directly from those images. Doing this in a lab, he said, typically costs around €100 a sample. “Instead of paying that for every single sample, you can just use the image that you have in any case and try to predict the gene expression out of that.”
Right now, the team is focused pretty much exclusively on computational pathology of tissue sample images, primarily through hematoxylin and eosin (H&E) stains that reveal the structure and morphology of the tissue sample. These stains are also used by clinicians and labs for diagnosis. Cancer is by far the most common application of this kind of analysis.
Obstacles & opportunities
Of course, working with health data is tricky, even several years after the advent of Covid. “That’s an issue at the moment,” Ruusuvuori said when asked about access to clinical data. “We can’t really have the clinical samples from the clinics that are used for treating the patients, so at the moment what we are using is the public datasets because of the very strict regulations in Finland … which we are all the time trying to change.”
“LUMI, at the moment, [being] in Finland, it cannot be used for processing these routine clinical samples,” he added. “So that’s a major obstacle for us to achieve what we really want to achieve.”
Until those obstacles change, the team is using tens of thousands of public domain tissue sample images to train their model, which the team built from the ground up.
“We are pushing as hard as possible to be one of the first ones who will build the very large pre-trained models to understand histology,” Ruusuvuori said. “I think we have a good chance to really build a model that is more comprehensive than what has been available before.” He elaborated that when working with prostate cancer data with colleagues from Sweden – before they had access to LUMI – the model was able to perform cancer detection and grading at expert-level accuracy. “Now we can really try to scale it up with much, much more data using LUMI.”
On LUMI, the team is using the LUMI-G partition – the most powerful segment of the machine, comprising the vast bulk of its Linpack rating. The jobs run by ComPatAI each take between 50 and 100 of LUMI-G’s 2,560 nodes.
While ComPatAI’s team is a research group, they do have an eye toward commercialization. “Research can only take you so far,” Ruusuvuori said. “We really want to bring this technology to the point of benefit of the patient.”
The research team is currently applying for more funding and time on LUMI, and they expect to share preliminary results from their initial research on the new supercomputer (maybe) later this year – but Ruusuvuori said they are “hoping to continue this research for a long, long time.”
“We are not pathologists or medical doctors,” he added, “but we try to solve real-world problems and try to help pathologists and physicians cope with the flood of data that they get.”