How Auria Biobank and Silo AI use computer vision to find new information to fight cancer

Auria Biobank (Turku University Hospital and University of Turku) and Nordic AI service and solution provider Silo AI announce today their collaboration in investigating computer vision-powered solutions for novel cancer diagnostics in digital pathology. The research project focuses on analyzing digital pathology images of tissue samples from one of the most common skin cancers. Silo AI’s AI Scientists leverage computer vision in order to analyze the vast dataset of these digital pathology images. 

Research with the goal of understanding if the cancer will spread

As part of Turku University Hospital, Auria Biobank, is, among other things, a repository of over 1M+ tissue samples from patients. Around 10 similar biobanks in Finland use their samples to contribute to medical research aimed at developing new medications and forms of treatment. In Auria, most of the tissue samples are from different cancers.

Together with Silo AI, Auria set out to investigate whether a certain type of skin cancer will spread to other parts of the body in the future. In scientific terms, it’s about understanding whether the tumor will metastasize or not. The computer vision-based classification is done based on the first digitized tissue sample resected from the patient’s tumor at the time of cancer onset. So far the clinicians have not found reliable markers that would predict later metastasization in the particular cancer under study.

“The skin cancer, cutaneous squamous cell carcinoma, we’re investigating is one of the most common cancers. By leveraging the latest computer vision technologies, our aim is to find significant new information that supports clinicians in their diagnosis work in assessing the risk of metastasis. This is expected to have an important impact on the patient outcome”, says Veli-Matti Kähäri, Professor of Dermatology at Turku University and Chief Physician at The Department of Dermatology of Turku University Hospital.

Kähäri is leading the research group that collaborates with experts from both Silo AI, Auria and Turku University.

The research interest rises from a research paper published in Nature (2018), that investigated lung carcinoma tissue. The paper suggested that determining the tumors’ driving oncogenes might be possible from digitized samples. In a similar vein, the hypothesis is that there might be previously unidentified histopathological differences between those cancer cells that will spread, and those that won’t – and that this could be seen in advance. 

The paper gained significant attention already at the time of publishing, and has been in Auria’s interest ever since, according to Auria’s Director, Lila Kallio:

“We started to investigate the possibilities at Auria for a similar approach immediately as these scientific results were published. For example, we have hundreds of skin cancer samples which can be grouped based on the clinical data collected by the collaborating clinicians at Turku University Hospital, and with the help of Silo AI, we believe we can discover novel significant findings in an efficient way.”

Mikko Tukiainen, AI Scientist at Silo AI, dives deeper into the project: “In short, we’re studying if there exists morphological differences that a computer would be able to see. Human pathologists have so far been unable to identify such features in the cancer in question. For the patients, having the information on whether or not the cancer will become metastatic can be crucial both in terms of on-going treatments but also in understanding the need for follow-up monitoring.”

Explaining AI decision-making helps human pathologists learn from computer vision-based analysis

Understanding the indications of future metastasis will eventually help human pathologists to learn from what the computer vision sees in the cancer tissue. Computer vision permits us to carefully analyze hundreds of images of cancerous tissue, and it is possible that the AI solution is able to find something that humans have so far been unable to see. Therefore, another core element of the project is to be able to explain and show what the AI solution is basing its decisions on. 

For this purpose, Silo AI’s AI Scientists have been creating various decision heatmaps with techniques such as Grad-CAM. With these explanation methods, clinicians are better enabled to unravel what the AI solution sees, as its decision making process may not be intuitive to natural human thinking at first.

By visualizing explanations with Grad-CAM, scientists are able to understand which parts of the image are most affecting the AI decision-making. Here, Grad-CAM highlights the regions that the AI technology judged as tumorous. The image is not related to the Auria project, but is part of a Kaggle challenge that Silo AI’s AI Engineer Joni Juvonen took. Mikko and Joni have taken several Kaggle challenges together to improve cancer tissue analysis with computer vision. 

Vast datasets go all the way down to cellular level

Computer vision-powered analysis is not a trivial task: digital pathology tissue images are extremely detailed as they cover all the way down to the cellular level of the patient. Therefore one single image can be several gigabytes in size, and needs to be treated in smaller pieces.

In addition, the data quality varies because of the different tissue sample staining protocols and scanners used in the different laboratories. Knowing these differences and training the AI model accordingly requires expertise both on the medical and clinical side, and on applying computer vision for human tissue analysis. Mikko, who leads the project at Silo AI, has several years of experience on medical computer vision, and has also published a paper on cancer related research in Finland’s largest medical journal Duodecim.

“I’m passionate about using AI for good. It is extremely interesting to see whether AI technologies can find any indication for a given skin cancer case later becoming metastatic, particularly as currently there are no known clinical, genetical or histopathologic cues for this. Together with Auria, however, I’m confident that we have a good opportunity to leverage both organizations’ expertise and data to find out more.”, Mikko concludes.

The images in this blog post are not related to Auria.