Machines working for your well-being
Artificial intelligence (AI) can produce better analysis of medical images.Thursday, September 12, 2019
Improving medical diagnoses, extending autonomy, preventing suicides: machines equipped with artificial intelligence are being touted as the key to revolutionizing the health sector.
Interpreting medical images is an art. The proof: it can take a highly trained radiologist up to a week to analyze and then interpret a handful of 3D digital brain scans for a single patient. Christian Desrosiers, a Professor at École de technologie supérieure (ÉTS) in Montréal, points out that, when it comes to identifying tumours, which can appear in highly diverse forms, the process can take much longer. He explains: “If you consider the relatively high average salary among radiologists, along with the volume of analyses to be carried out, it is easy to see that there is room for improvement”.
Artificial intelligence (AI) can produce better analysis of medical images. When given access to patients’ scans, AI machines can tell their story. Better still, according to Christian Desrosiers, they are able to discern certain subtleties that escape the eye of even the most experienced clinicians: “Unlike humans, computers can detect patterns within a series of images that may at first seem unrelated”. However, we must exercise prudence, because this technology is far from proven. The day when algorithms assist radiologists, or even replace them, has not yet arrived.
They still have to demonstrate their potential. Toward this end, and within the context of a project being conducted in collaboration with researchers from the McGill University Health Centre, AI developed by Christian Desrosiers’ team and “imprisoned” within an application has been able to produce accurate results in 80% of cases for patients suffering from brain cancer. In order to do so, it identifies predictive markers (the dimensions, location and size of the tumour, etc.) contained in hundreds of medical images that have been previously annotated by hand by clinicians. Desrosiers explains: “The algorithm was able to identify the disease, determine the stage, and pinpoint whether the patient was situated below or above the average survival rate”.
Thousands of scans would be required to attain an accuracy rate approaching 100%, but datasets of that type are not currently available in Québec. Professor Desrosiers clarifies: “It’s not that there aren’t enough medical images to analyze, but that they are not necessarily annotated”. Nevertheless, he plans to take advantage of extensive open databases, such as the UK Biobank, to resolve this problem.
Translation of an excerpt from Québec Science