Artificial intelligence (AI) systems can diagnose prostate cancer biopsies with the same level of accuracy as specialist uropathologists, and better than many general pathologists, a new research has found.
In the study published in the science journal Nature Medicine, the performance of different AI models was tested against the gold standard diagnosis of prostate cancer.
The researchers organised a global competition to build AI models to diagnose more than 10,000 prostate biopsies.
Over 1,000 AI developers from 65 countries participated in the competition, sending in 1,010 algorithms to be assessed for accuracy in diagnosis, making it the largest competition to be held into the use of AI in pathology.
Fifteen of the algorithms were selected to have their performance measured against diagnoses made by specialist uropathologists and general pathologists.
The research provides the first independent evaluation of AI algorithms across different patient populations and pathology labs and across reference standards developed by expert panels of uropathologists from the US and Europe, said pathologist professor Brett Delahunt from the Department of Pathology and Molecular Medicine at the University of Otago, Wellington.
Prostate cancer is the second most commonly occurring cancer in men and the fourth most commonly occurring cancer overall. In 2020, it accounted for more than 14 lakh cases and over 3 lakh deaths globally.
Delahunt said achieving more precise diagnoses is key to reducing the number of deaths.
“Assessment of biopsies is crucial when it comes to making decisions on prostate cancer treatment – but there can be significant variations in the assessments made by different pathologists.
“Standardised AI models could really make a difference when it comes to improving outcomes for this disease,” he noted.
Pathologists characterise tumours into different ‘Gleason’ growth patterns, with biopsy specimens categorised into one of five International Society of Urological Pathology grade groups.
But the process is quite subjective and can lead to both ‘undergrading’ and ‘overgrading’ of prostate cancer biopsies, Delahunt said.
On the other hand, because the algorithms would likely miss fewer cancers than the pathologists did, AI could be used to reduce the workload of pathologists by automating the identification and exclusion of most benign biopsies, he noted.
