Developing AI-based software to identify prostate cancer in MR images.

Highlights 2020

The project's contribution to the Centre for Digital Life Norway annual report 2020.

The purpose of the project is to develop artificial intelligence (AI)-based software to support the detection and characterisation of prostate cancer on MR images. One highlight in 2020 was the extensive database of MRI scans and clinical information from Norway, the Netherlands and Taiwan (>3000 patients) now made available to the project at HUNT Cloud, a secure server adapted for storage and computational analysis of sensitive data.

During 2020, our research focused on development of analytical pipelines for the project. We have optimised deep learning-based routines for automatic segmentation of the prostate and further developed a novel quality control system designed to capture when segmentation fails, a situation which could lead to detrimental errors in the final decision support. Whereas the multicentre and multivendor nature of our data provides unique advantages, there is also a major, accompanying challenge of intensity variability. Hence, image normalisation is a required part of the pipeline. For this, we have developed a novel method for automated dual-reference tissue normalisation. We also proved that these steps are important for optimising novel Generative Adversarial Networks to detect prostate cancer in the images. Finally, active surveillance is a management strategy in prostate cancer involving longitudinal monitoring to detect disease progression in patients with low-risk disease, with the aim of deferring treatment. We now develop a framework based on MRI features derived from repeated scans to support the detection of disease progression.

For us, the added value of being part of a transdisciplinary centre is the available expertise from life sciences, data sciences and social sciences, as well as the possibility to directly explore the relevance of this expertise through a direct link with the clinic. This is important to our project as the problems to be solved evolve from real-world clinical limitations, are complex in nature and require out-of-the-box thinking.

A T2-weighted MR image of the prostate gland (mid-slice) of a patient with biopsy-confirmed prostate cancer. The gold standard radiology manual segmentation is labelled with colours: tumour (blue), peripheral zone (red) and the remaining prostate zones (green). Credit: Mohammed Sunoqrot/NTNU.


Event with support from the Centre for Digital Life Norway 2020: 26 November, Trondheim, digital: Artificial intelligence in cancer diagnostics – between hope, realism, and ethical challenges

Scientific publications 2020: 2

Project overview

Project lead: Tone Frost Bathen
Institution: NTNU
Partners: St. Olav's Hospital and Norwegian University of Life Sciences (NMBU)
Duration: 2019–2022

Research group

1 in 8 men will have prostate cancer in their lives. PROVIZ is developing AI-based software to correctly identify prostate cancer in MR images, reducing the time and cost of diagnosis.

Norway was the first country to implement an integrated cancer care pathway that uses multi-parametric magnetic resonance imaging (mpMRI) as the first diagnostic tool for men with suspected prostate cancer based on elevated levels of prostate specific antigen (PSA) in their blood. However, the PSA blood test has a high false-positive rate, leading many men to get unnecessary mpMRI’s. The popularity of this blood test has created an excess of images for radiologists to analyze.

PROVIZ’s research will support radiologists by expediting the process of detecting cancer in the mpMRI images and by guiding biopsy targets. Today, radiologists rely on their training and experience to identify whether cancer is present and to guide biopsies to further test cancer aggressiveness. PROVIZ’s AI analyzes the quantitative information in mpMRI images (anatomy, vascularization, and cellularity) and builds 3D models to identify and visualize potential cancer risk, reducing the burden on radiologists and lowering biopsy risks. 

The software PROVIZ is developing is based on a unique Norwegian dataset of MRI and clinical information from over 1600 patients with and without cancer, as well as collaboration with international teams collecting data in The Netherlands and Taiwan. This large data set allows the team to account for variation between clinical sites and create a truly universal diagnostic platform.  

Clinicians, patients, and research participants have been involved in development from the start of the project to ensure the research aligns with stakeholder and social needs. This inclusive team has the background to anticipate both positive and negative outcomes of their work. In addition to experts in MRI technology, information analysis, radiology, oncology and communication between researchers and clinicians, the PROVIZ team includes a doctoral candidate specifically focused on ensuring that responsible research and innovation is integrated into the project. 

Adaptable AI-based software has the potential to reduce health care costs, ease the burden on medical personnel, and obtain better treatment outcomes. Once the project is complete, PROVIZ aims to share their data set and invite others to use it to educate the field and develop additional tools. 

PROVIZ is headed by Tone Bathen at the Norwegian University of Science and Technology in Trondheim. This lab is funded by the Research Council of Norway and is one of the multidisciplinary research projects within Centre for Digital Life Norway.

By Matthew Davidson