AI can be used to detect and grade prostate cancer
“Our results show that it is possible to train an AI system to detect and grade prostate cancer on the same level as leading experts,” says Pekka Ruusuvuori, leader of the Bioimage informatics research group at the Faculty of Medicine and Health Technology at Tampere University, who led the project on the Finnish side.
“This could significantly reduce the workload of uro-pathologists and allow them to focus on the most difficult cases. Our study shows that computational tools have the potential to benefit routine diagnostics in pathology.”
To train and test the AI system, the researchers scanned more than 8,000 biopsies taken from some 1,200 Swedish men aged 50–69 years. About 6,600 of the samples were used to train the AI system to spot the difference between biopsies with or without cancer.
The remaining samples, and two additional sets of samples collected from other databases, were used to test the AI system. Its results were compared against the assessments of 23 world-leading uro-pathologists.
The findings showed that the AI system was near-perfect in determining whether a sample contained cancer or not. The system was also capable of estimating the length of the tumor in each biopsy, an important piece of diagnostic information, with high precision. When it comes to determining the severity of the prostate cancer, the so called Gleason score, the AI system was on par with the international experts.
“The idea is not that AI should replace the human involvement, but rather act as a safety net to ensure that pathologists don’t miss some cancers,” says Martin Eklund, associate professor at the Department of Medical Epidemiology and Biostatistics at Karolinska Institutet, who led the study.
“It could also serve as an alternative in parts of the world where there is a complete lack of pathological expertise today.”
A computational challenge
While the large amount of data available for the study was a key for success, it also represented a computational challenge. Training an AI system consists in feeding the algorithm with countless examples, in this case millions of digital images extracted from the scanned biopsies. In practice, the associated computation is only feasible on graphics processing units. The state-of-the-art resources provided by TCSC - Tampere Center for Scientific Computing and CSC - IT Center for Science were an enabling factor for the study.
“Being able to train the final AI model on a dataset of this scale within days is one thing – but perhaps even more valuable has been the fast-paced testing of different ideas during the earlier stages of the study. By experimenting with many alternative approaches in parallel, we learn quickly,” says Kimmo Kartasalo, doctoral student at Tampere University and joint first author of the article.
“Utilizing the hardware efficiently does not happen by the push of a button, though. Figuring out how to process the histopathological image data while avoiding computational bottlenecks formed a considerable part of the whole project effort.”
Clinical use will require a focus on quality
The initial results are promising, but there is more work to be done before AI-based diagnostics may be rolled out in everyday clinical practice, according to the researchers. A key prerequisite for widespread adoption of AI in the clinic is ensuring that the system can function reliably in a real-world setting, where biopsies prepared in different laboratories and scanned on different types of digital scanners typically differ in their appearance.
“Ensuring that the AI system can generalize across different data sources and produce reliable diagnoses even when facing something unexpected in the data is crucial,” says Pekka Ruusuvuori.
“Now that we have shown that we can reach very high performance in a realistic but controlled setting, the next step is to focus on the potential error sources that the AI can be exposed to in broader clinical use.”
● Prostate cancer is the most common cancer and the second most common cause of cancer death in men in Finland.
● More than 20 million prostate biopsies are examined every year in Europe and in the U.S.
● There is a shortage of pathologists in Western countries and internationally, and in many developing countries, the ratio is fewer than one pathologist per 1 million people.
● Efforts to reduce mortality in prostate cancer are often hampered by the difficulty in making objective and reproducible pathological assessments of prostate biopsies.
Funding: Academy of Finland, Cancer Society of Finland, Emil Aaltonen Foundation, European Research Council, Finnish Foundation for Technology Promotion, The Finnish Society of Information Technology and Electronics, Industrial Research Fund of Tampere University of Technology, KAUTE Foundation, Orion Research Foundation, Svenska Tekniska Vetenskapsakademien i Finland, Swedish Cancer Society, Swedish eScience Research Center, Swedish Research Council, Swedish Research Council for Health, Working Life, and Welfare (FORTE), Tampere University Foundation, Tampere University graduate school, TUT on World Tour, Tutkijat maailmalle, Walter Ahlström Foundation.
Publication: “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
Pekka Ruusuvuori, Assistant professor, Institute of Biomedicine, University of Turku &
Adjunct professor, Faculty of Medicine and Health Technology, Tampere University
Phone: +358 50 318 2407
Email: email@example.com, firstname.lastname@example.org
Photograph: Kimmo Kartasalo