An estimated 1.41 million people lost their lives to prostate cancer in 2020 according to the World Health Organisation, making it the fifth leading cancer killer after breast, lung, colon, and rectal cancers. But there’s a big problem – overdiagnosis and overtreatment.
One in nine men will develop the condition at some point in their lives, and while most will recover with treatment, many don’t require any intervention at all. This is because prostate cancers grow slowly and may not pose any health risk to them.
To get around this problem, researchers globally are increasingly turning to machine learning and artificial intelligence to help in grading biopsies of prostate cancer. They hope this can not only improve on the accuracy of diagnosis, but also solve the issue of a shortage of pathologists.
Currently a technique known as the Gleason grading system is used to assesses the aggressiveness of the cancer cells’ growth. But this is done manually and is highly subjective, leading to differences in opinion between pathologists for the same screening.
Ultimately, this can mean someone who doesn’t require treatment – either surgery or radiology – may end up getting it, while others who might require treatment may not. According to the American Cancer Society, the major downside of overtreatment of prostate cancer is that it can affect quality of a man’s life as it can lead to urinary, bowel and/or sexual side effects.
AI for accuracy
“Because these biopsies are small samples of tissue, they only provide a small glimpse into the larger picture,” says Cameron Chen, a software engineer at Google Health, who was part of a challenge that the company ran to speed up the development and training of AI algorithms to diagnose prostate cancer.
“This is a key part of the data we rely on to decide whether or not to remove someone’s prostate or pursue other treatment decisions such as radiation,” adds Chen.
Google Health, a division of Alphabet that’s working on bringing AI in diagnosis of cancer and to prevent blindness, in April 2020 organised the Prostate cANcer graDe Assessment (PANDA) Challenge, which saw the participation of over 1,000 teams from 65 countries.
The results of the challenge, which were published earlier this year in Nature, showed significant gains in the accuracy and time taken to develop such algorithms.
“Impressively, the first team achieving high agreement with the prostate pathologists at above 0.90 on the internal validation set occurred within the first 10 days of the competition,” Google Health wrote in a blogpost. “By the 33rd day, the median performance of all teams exceeded a score of 0.85.”
Chen told Happiest Health that in a separate study conducted to test the usefulness of such artificial intelligence systems, they found that the tools increased the accuracy as well as efficiency of pathologists.
Coming sooner than you think
The US Food and Drug Administration (FDA) authorised the first software, built by Paige Inc, to identify prostate cancer in scans in September last year. The government body said that it had reviewed a study that found Paige Prostate’s AI software improved detection of prostate cancer by 7.3% compared to pathologists’ doing so manually using existing techniques.
“Identifying areas of concern on the biopsy image can help pathologists make a diagnosis that informs the appropriate treatment,” Tim Stenzel, director of the FDA’s Office of In Vitro Diagnostics and Radiological Health, said in the agency’s announcement. “The authorization of this AI-based software can help increase the number of identified prostate biopsy samples with cancerous tissue, which can ultimately save lives,” FDA’s Stenzel added.
Researchers assessing artificial intelligence systems for diagnosing prostate cancer also say that they have the potential to become essential tools for urological radiologists, if not for the field of urology in general.
“In prostate cancer, the use of AI overall has shown to be beneficial to aid in a standardized pathological grading to assess prostate cancer stratification and treatment,” a group of researchers from Miami University’s Miller School of Medicine wrote in a paper reviewing the use of AI in prostate cancer.
Researchers are also hoping that AI-based tools will help pathologists make better diagnoses and eventually replace them, which will free up their time to do more work.
Training the AI
Verily, another Alphabet company, in partnership with digital pathology solutions firm Lumen entered a partnership to develop products that can “diagnose, prognose, and guide prostate cancer therapy selection with improved objectivity and efficiency.” The algorithms used by Verily were developed by Google Health.
The Alphabet company will validate its algorithms for accuracy with Lumen’s database of prostate cancer biopsies. This highlights the biggest impediment in the development of AI tools in diagnosis – the availability of medical diagnosis data, including access to imaging scans that have been graded by specialists.
In part, this is one of the things that Google Health was aiming to solve with its PANDA challenge. It made available datasets from the US and EU, to the over 1,290 developers that took part, and even to others who would want to continue research in the space.
“In the context of the challenge, the AI algorithms developed by the teams improved rapidly over the course of the challenge. The curated datasets and availability of high-quality AI development software tools contributed to this velocity,” Chen said, adding that a greater diversity of data sources will ensure that AI models will be applicable to a wider population.
