PAIGE, developing AI software for pathology, received $100M in funding in early 2021 and FDA de novo clearance for prostate cancer AI in September 2021.
In this blog, I recap some of the key timepoints, provide links, compare to AI in radiology, and summarize a few remarks on clinical performance from yesterday's PAIGE public webinar.
It was big news in January 2021 when PAIGE landed a $100M Series C funding round (here, here.) There was even more excitement on September 22, 2021, when PAIGE announced de novo marketing authorization for Paige Prostate "a clinical grade AI solution for prostate cancer detection." FDA marked the day with a press release (here).
PAIGE @ FDA
The FDA website provides a special page for de novo designations here, where you can find the Paige Prostate classification order (DEN200080 here). At some point, this FDA page will also display a link to a detailed clinical trial safety and effectiveness report (typically 20-40 pages). But the appearance of the safety, effectiveness, and clinical trial report can lag the marketing decision by a few weeks or months.
The new (aka "de novo") classification group is 21 CFR 864.3750, "Software algorithm to assist users in digital pathology."
Note that the products are intended to "assist users" not "diagnose" the slides. This is very similar to FDA regulation of AI and machine learning in radiology, where the software is allowed to flag parts of images or to prioritize cases within a queue so that patients with health-critical lesions get their diagnoses before patients with normal scans.
The Paige Prostate FDA product code is QPN (same title as 364.3750) and is here. I don't think the regulation for category 864.3750 has been released yet. FDA assigns numbers in advance of finalizing the public text of the classification.
(In contrast, the paperwork for the 2020 Viz-AI software is complete. Viz-AI has FDA approved radiology software for its ContaCT product, and you can find the full 18pp safety & effectiveness decision summary online at FDA (DEN 170073). And, the public regulatory classification into which ContaCT falls is for "radiological computer aided triage and notification software." This was published 4/1/2020 as 21CFR 892.2080 here. In contrast, for Paige Prostate, neither the FDA Decision Summary nor regulatory classification .3750 are online yet.)
The PAIGE indications use state:
Paige Prostate is a software only device intended to assist pathologists in the detection of foci that are suspicious for cancer during the review of scanned whole slide images (WSI) from prostate needle biopsies prepared from hematoxylin & eosin (H&E) stained formalin-fixed paraffin embedded (FFPE) tissue.
After initial diagnostic review of the WSI by the pathologist, if Paige Prostate detects tissue morphology suspicious for cancer, it provides coordinates (X,Y) on a single location on the image with the highest likelihood of having cancer for further review by the pathologist. Paige Prostate is intended to be used with slide images digitized with Philips Ultra Fast Scanner and visualized with Paige FullFocus WSI viewing software.
Paige Prostate is an adjunctive computer-assisted methodology and its output should not be used as the primary diagnosis. Pathologists should only use Paige Prostate in conjunction with their complete standard of care evaluation of the slide image.
On its website, PAIGE notes that "Paige Prostate has been tested on slides from more than 200 institutions and in multiple peer-reviewed studies," but I didn't easily find a publications list (see [*]). The FDA S&E data may pop up online and show much of the data prior to some of the the journal publications).
I did have a couple recent prostate AI publications at hand. Last year, Lancet published a detailed paper by a Swedish group, Strom et al., on "artificial intelligence for diagnosis and grading of prostate cancer in biopsies," here. See also a Lancet Digital Health paper Pantanowitz et al., "artificial intelligence algorithm for prostate cancer diagnosis on slides," here, reflecting software being commercialized by IBEX. (And see footnote [*]).
This Week's Webinar
In a public webinar on October 26, PAIGE discussed its journey through the FDA and also discussed some very interesting performance data for its product, describing a key publication as being submitted. PAIGE described a study with samples from 5 countries, 22 states, and 217 institutions, and review by 16 pathologist (some GU specialists, some generalists). While the FDA labeling describes the software as flagging either 0 or 1 lesion for further pathologist review, the webinar emphasized that in a protocol, slides could be signed out as no cancer, or as cancer, or as flagged as indeterminate for further review ("deferral," with other pathologists and/or special stains). False negatives for generalists dropped from 12% to 4% for generalists, and from 9% to 5% for urologic specialists. This improves the sensitivity of diagnosis (false negatives mean missed cases mean lack of sensitivity.) However, specificity also improved. False positives fell from circa 2-3% to circa 1-2% for both generalists and specialists.
PAIGE summarized that there was an decrease of 24% in "unnecessary deferrals" and an increase of 59% in "necessary deferrals." An unnecessary deferral would be a slide flagged initially for deferral, but signed out as positive when the pathologist considered both his diagnosis and what was flagged by Paige. A "necessary" deferral is when a slide was pulled from the "diagnosis benign" queue, and brought into further review that led to a correct cancer diagnosis (for example, with conferral or with immunohistochemistry).
In the webinar, PAIGE listed key related FDA approvals as the Phillips IntelliSite solution in 4/2017, the Aperio solution in 5/2019, the Sectra pathology module in 4/2020, the Paige FullFocus system in 7/2020, and the Paige Prostate system in 9/2021.
While AMA CPT and CMS don't have clear payment pathways yet for AI in slide pathology, CMS has provided add-on payments (what are called NTAP) to radiology AI products used as part of inpatient care, with several products approved successively by CMS in August 2020 and then August 2021 in annual inpatient rulemaking. See my reference above to 21CFR 892.2080.
(*) Some additional publications
Some other interesting publications in the field include:
Mata, 2021, PMID 34597215, on AI assisted diagnosis viewing MRI and biopsy as a collective diagnostic process.
Purysko, 2021, PMID 34597239, Op Ed on the above.
Perincheri, 2021, PMID 33782551, Paige Prostate tested at Yale.
da Silva, 2021, PMID 33904171, "real world application of Paige AI prostate cancer detection."
Raciti, 2021, PMID 32393768, Paige AI system increases detection of prostate cancer. Writing, "Without Paige Prostate Alpha, pathologists had an average sensitivity of 74% and an average specificity of 97%. With Paige Prostate Alpha, the average sensitivity for pathologists significantly increased to 90% with no statistically significant change in specificity. With Paige Prostate Alpha, pathologists more often correctly classified smaller, lower grade tumors, and spent less time analyzing each WSI."
For an earlier paper on development of the Paige AI system with weakly supervised machine learning, see Campanella, 2019, PMID 31308507.
Though not directly reporting on AI accuracy, an interesting paper is "integrated digital pathology at scale," by Schuffler, 2021, PMID 34260720. Writing, " We built a vendor-agnostic, integrated viewer for reviewing, annotating, sharing, and quality assurance of digital slides...used by 926 pathologists and researchers evaluating 288,903 slides...An interconnected Honest Broker for Bioinformatics Technology (HoBBIT) to systematically compile and share large scale digital pathology datasets...We highlight major challenges and lessons learned." (It sounds like Google Earth for pathology.)
Addendum. Although not directly related to PAIGE, another item related to advanced pathology. See Balaur et al. in Nature, 6 October 2021, "Colorimetric histology using plasmonically active microscope slights," aka, "the slide becomes a sensor." Nature here. Trade press here and here. Youtube here.
Earlier study by Powley on patient derived explants for anti-cancer drug discovery here.