null Artificial intelligence detects and grades prostate cancer nearly without error

Researchers digitally scanned more than 8,000 prostate biopsies to train and test the artificial intelligence. Photo: Kimmo Kartasalo.

Artificial intelligence detects and grades prostate cancer nearly without error

Researchers from the University of Tampere and Karolinska Institute in Stockholm trained artificial intelligence to diagnose and grade prostate cancer. The artificial intelligence system could correctly identify biopsies containing cancer nearly without error. The supercomputer Puhti-AI, which was launched on a pilot basis in the early autumn of 2019, contributed to calculating the final phase of the work. The study was published in the prestigious "The Lancet Oncology" magazine.

Researchers digitally scanned more than 8,000 prostate biopsies to train and test the artificial intelligence. An artificial intelligence system consisting of ten deep neural networks was trained to distinguish between benign and cancerous prostate biopsies. Using the so-called Gleason Score, the cancerous tissue was further classified into categories describing the aggressiveness of the cancer.

Approximately six million images extracted from digitally scanned biopsies were used to train the artificial intelligence. There were approximately 30 terabytes of image data, and the final training data consisted of more than 2 trillion pixels.


"Using Puhti went almost astonishingly well during the pilot phase, and the increase in performance was immediate"





– Neurogrid models are generally adapted using GPUs, which, at best, achieves a speed tens of times faster than CPU computing. Roughly speaking, with a normal single CPU, it would have taken at least several months if not years to train our model. Using the Taito GPU and the Narvi cluster in Tampere, we were able to train the model in 1-2 weeks. In the final stages of the work, the model had to be re-adapted once more, but, fortunately, at this stage we were able to use Puhti, which dropped the total computing time to 2-3 days, explains Pekka Ruusuvuori. Ruusuvuori was responsible for the University of Tampere contribution to the research project.

–  Using Puhti went almost astonishingly well during the pilot phase, and the increase in performance was immediate, says Kimmo Kartasalo, a doctoral researcher at the University of Tampere, who was responsible for the practical implementation of the research together with Swedish doctoral researchers.

GPU Processors and high-speed disks play a key role

In addition to graphics processors, high-speed disk capacity was found to play a key role in the computing. Due to the extensive scope of the data, it is not possible or efficient to read it in its entirety in the central memory, so it must be possible to buffer image data to the central memory on the fly as it is processed by the GPU.

–  Optimizing this process was essential to our work, and we will continue it in our subsequent projects, explains Ruusuvuori.

CSC's computing capacity was critical to the research, and working with such extensive data would not have been possible without state-of-the-art GPU capacity.

–  Part of our role in this collaborative effort was to facilitate the making of extensive computing. In addition to training the final model, computing resources also played a key role in the early stages of the work, as it then allowed us to prototype, for example, different neural network architectures and different parameter combinations in parallel using the Taito GPU and Tampere cluster. In this case, we soon figured out which direction the development work should take. With Puhti, this kind of prototyping will be much faster, thus making our research work more efficient, adds Ruusuvuori.

Near-perfect detection accuracy

The results showed that the AI system was able to distinguish between cancerous biopsies and benign biopsies almost without error. Artificial intelligence also succeeded in assessing with a high degree of accuracy the length of cancer tissue in each biopsy, which is essential additional information from a diagnostic standpoint.

When analyzing how artificial intelligence performs in assessing the aggressiveness of prostate cancer using the so-called Gleason Score, the system was also found to produce comparable results as those obtained by experts.

–  However, the goal is not to replace human experts, but rather to provide pathologists with a tool that can, on one hand, improve work efficiency, while also promoting patient safety by acting as a safety mechanism, explains head of the study, Associate Professor Martin Eklund of the Karolinska Institute.

Even though the results have been promising, more work is still needed before Ai-based diagnostics becomes commonplace in clinical use. A key prerequisite for the extensive use of artificial intelligence is the assurance that the system will also function reliably in clinical practice, where the appearance of biopsies varies, for example, between different laboratories and different scanning equipment.

–  We have to ensure that the high level of performance is standardized in data collected from different sources and that artificial intelligence can function reliably even when it encounters something unexpected in the data. In a realistic. but controlled experiment design, it is possible to achieve extremely high diagnostic accuracy, and the next step is to focus on the causes of errors, which artificial intelligence must be able to tolerate in extensive clinical use, says Ruusuvuori.


University of Tampere press release: AI can be used to detect and grade prostate cancer



"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

Published originally 31.1.2020.

More about this topic » Go to insights and news »

Tommi Kutilainen