FDA’s authorisation of AI image classifier for prostate cancer

Connor Haile writes:

The agency granted its first clearance for a cancer diagnosis AI program to Paige, a New York-based company launched in 2018 with data and digital pathology tech from Memorial Sloan Kettering Cancer Center. The product, Paige Prostate, analyzes slides of biopsied prostate tissue to spot the hallmarks of malignant cells. The software highlights areas of a standard prostate biopsy image that are most likely to contain cancer and flags them for further review by a trained professional.

A press release from FDA has more details

The FDA evaluated data from a clinical study where 16 pathologists examined 527 slide images of prostate biopsies (171 cancer and 356 benign) that were digitized using a scanner. For each slide image, each pathologist completed two assessments, one without Paige Prostate’s assistance (unassisted read) and one with Paige Prostate’s assistance (assisted read). While the clinical study did not evaluate the impact on final patient diagnosis, which is typically based on multiple biopsies, the study found that Paige Prostate improved detection of cancer on individual slide images by 7.3% on average when compared to pathologists’ unassisted reads for whole slide images of individual biopsies, with no impact on the read of benign slide images.

They reviewed the software through “DenoVo classification” system

The De Novo process provides a pathway to classify novel medical devices for which general controls alone, or general and special controls, provide reasonable assurance of safety and effectiveness for the intended use, but for which there is no legally marketed predicate device. De Novo classification is a risk-based classification process.

I couldn’t find any references to how the software added more complexity in the workflows. I remember my attempt to get digital pathology in my institution- they were not interested. I wanted to combine the digital path with genomics, but these “innovations” were only on the fringes. It would require a considerable overhaul of the workflows, including grossing, labelling, analysis, and a substantial increase in manpower/costs. Besides, Paige didn’t mention the rate of false positives- which would accrue more costs.

Of course, I refrain from linking to press releases – they don’t mention the costs involved in cloud-based processing or data-localisation, and how images will be stored for legal requests/analyses. It opens up a whole can of worms if you start questioning. I am sure they will talk about it in any of the upcoming GU conferences with flashy pamphlets.