Endoscopy and pathology
AI that flags possible polyps during colonoscopy, or suspicious tissue on digitized biopsy slides, for the physician to review.
Computer-aided detection (CADe) systems such as GI Genius and EndoScreener flag possible polyps on the colonoscopy video in real time for the endoscopist to confirm. A 2023 meta-analysis of 21 randomized trials (18,232 patients) found an adenoma detection rate of 44.0% with CADe versus 35.9% without, but no increase in detection of advanced adenomas and more removal of non-neoplastic polyps. A pragmatic implementation trial reported that CADe did not improve colonoscopists' performance. In a deskilling study at four centers in Poland that used another system, Olympus ENDO-AID, the adenoma detection rate in colonoscopies done without AI fell from 28.4% to 22.4% after AI was introduced.
In pathology, FDA authorized Paige Prostate in September 2021 as the first AI-based software designed to identify an area of interest on prostate biopsy images; in the supporting reader study, the software raised pathologists' average sensitivity for cancer on individual slide images by 7.3 percentage points, FDA said. Open questions include effects on patient-important outcomes, which the meta-analysis called unclear, and on endoscopists' own skills.
EndoScreener
EndoScreener is stand-alone software for real-time automatic detection of polyps in the colonoscopy video stream during the procedure, and the physician must confirm its findings, according to the company's 510(k) summary.
Why this grade. A randomized trial of 790 patients at Sichuan Provincial People's Hospital in Chengdu, China (Liu et al., Therapeutic Advances in Gastroenterology, 2020), which names the system EndoScreener, found an adenoma detection rate of 29.01% with it versus 20.91% with routine colonoscopy, its primary outcome.
The evidence
In the 790-patient trial, which showed EndoScreener on the main monitor, adenomas per colonoscopy were 0.48 with the system versus 0.29 without, the added detections were mainly non-advanced diminutive adenomas, serrated adenomas and hyperplastic polyps, and endoscopist fatigue scores did not differ significantly between the groups. The company's 510(k) summary describes a U.S. randomized tandem colonoscopy study of 223 patients in which the adenoma miss rate was significantly lower, and first-pass adenomas per colonoscopy higher, when the system was used first. An earlier randomized trial by the same group (Wang et al., Gut, 2019) tested Shanghai Wision AI's polyp detection system in 1,058 patients at the same hospital and found an adenoma detection rate of 29.1% with the system versus 20.3% without, a gain driven by diminutive adenomas, with no significant difference in larger adenomas and more hyperplastic polyps found. That paper names the developer but not the product.
- FDA
- FDA cleared EndoScreener through 510(k) K211326 on Nov. 19, 2021, with GI Genius as the predicate device, according to the summary. The applicant was Chengdu Wision Medical Device Co.
- Limits
- The deciding trial was at one hospital in China, its added detections were mainly non-advanced diminutive adenomas, serrated adenomas and hyperplastic polyps, and Wision AI provided the system free of charge; a co-author disclosed consulting fees from the company.
GI Genius
GI Genius is a computer-assisted reading tool that helps endoscopists detect colonic mucosal lesions, such as polyps and adenomas, in real time during standard white-light endoscopy, according to Medtronic.
Why this grade. Two of three randomized trials of GI Genius met their primary outcome: COLO-DETECT, a pragmatic trial of 2,032 adults at 12 NHS hospitals in England (1.56 versus 1.21 adenomas per procedure), and a trial of 685 patients in Italy (adenoma detection rate 54.8% versus 40.4%); a trial in Spain found no difference in advanced neoplasia detection, its primary outcome.
Caution. In a before-and-after study at a large-volume center (Levy et al., American Journal of Gastroenterology, 2022), the adenoma detection rate (ADR) fell from 35.2% before GI Genius was introduced in all endoscopy suites to 30.3% after. The polyp detection rate also fell, from 40.9% to 36.5%.
The evidence
In COLO-DETECT, the ADR was 56.6% with GI Genius versus 48.4% with standard colonoscopy, and adverse events were similar. An earlier randomized trial of 685 patients at three centers in Italy (Repici et al., Gastroenterology, 2020) also found a higher ADR with the system, 54.8% versus 40.4%. A randomized trial of 3,213 patients with a positive fecal immunochemical test in Spain's colorectal cancer screening program (Mangas-Sanjuan et al., Annals of Internal Medicine, 2023) found no significant difference in its primary outcome, detection of advanced neoplasia (34.8% versus 34.6%), or in ADR, which was high in both groups (64.2% versus 62.0%).
- Used by
- It ran in routine colonoscopy at 12 NHS hospitals in England in the COLO-DETECT trial, and a large-volume center introduced it in all its endoscopy suites, Levy et al. reported.
- FDA
- FDA granted De Novo authorization DEN200055 on April 9, 2021, to the requester, Cosmo Artificial Intelligence, for a gastrointestinal lesion software detection system. A later 510(k), K261369, for the GI Genius Module 300 and ColonPRO US, was cleared on May 29, 2026, according to FDA's list of AI-enabled devices.
- Limits
- COLO-DETECT was open-label and funded by Medtronic, the trial with advanced neoplasia as its primary outcome found no gain, and the effect of computer-aided detection on patient-important outcomes remains unclear, a 2023 meta-analysis of 21 randomized trials concluded.
Ibex Galen
Galen Second Read, the FDA-cleared prostate product, raises case- and slide-level alerts with a heatmap when it detects tissue suspicious for prostate adenocarcinoma, according to its 510(k) summary. Ibex describes Ibex Breast as an AI assistant for breast pathology that automates analysis of hematoxylin and eosin slides and quantification of immunohistochemistry.
Why this grade. At Maccabi Healthcare Services in Israel, Galen Breast ran as a real-time second-read quality check on 5,954 routine cases, and its alerts led to the identification of cancers the reviewing pathologist had initially missed, a company-funded study reported (Sandbank et al., npj Breast Cancer, 2022).
The evidence
In a blinded external validation on 841 slides from 436 patients at Institut Curie in France and at Maccabi, reported in the same study, the breast algorithm had an area under the curve (AUC) of 0.99 for invasive carcinoma, with sensitivity of 95.51% and specificity of 93.57%, and an AUC of 0.98 for ductal carcinoma in situ. In routine use at Maccabi, Galen Breast raised invasive-cancer alerts on 363 slides (4.2%) from 272 cases initially diagnosed as benign. The evidence for Galen Second Read, the FDA-cleared prostate product, is a standalone study of 347 cases and a reader study in its 510(k) summary; in the reader study of 12 pathologists at four sites, combined sensitivity was 93.9% with the device versus 90.5% with standard of care, and combined specificity was 87.9% versus 91.1%. On that evidence alone it would be graded C.
- Used by
- Maccabi Healthcare Services' Institute of Pathology in Israel deployed Galen Breast, the product based on the study's algorithm, in December 2019, a company-funded study reported. Ibex lists deployments of Ibex Prostate at CorePlus in Puerto Rico and Betsi Cadwaladr University Health Board in Wales.
- FDA
- FDA cleared Galen Second Read, for prostate core needle biopsies, through 510(k) K241232 on Jan. 24, 2025, with Paige Prostate as the predicate. The grade rests on Galen Breast, which this clearance does not cover. Ibex's prostate page says Ibex Prostate is for research use only in the U.S. and not cleared by FDA; its breast page says the platform includes a solution that is FDA cleared.
- Limits
- Ibex funded the Maccabi study, two co-authors serve on its scientific advisory board, and the deployment report is descriptive, with no comparison group; in the prostate reader study, pathologists' specificity fell with the device.
Paige Prostate Detect
Paige Prostate Detect, authorized by FDA as Paige Prostate, helps pathologists detect foci suspicious for cancer while reviewing scanned whole-slide images of prostate needle biopsies stained with hematoxylin and eosin, according to Paige and FDA's decision summary.
Why this grade. In CONFIDENT P, a prospective trial in routine diagnostics that allocated 82 patients in alternating two-week blocks rather than at random (Flach et al., JCO Clinical Cancer Informatics, 2025), AI assistance lowered immunohistochemistry (IHC) use per detected prostate cancer (relative risk 0.55 at the patient level).
The evidence
In CONFIDENT P, IHC use per detected cancer also fell at the slide level (relative risk 0.41), and median assessment time per slide did not differ significantly (139 seconds with AI assistance versus 112 seconds without). An external validation at Yale on 1,876 core biopsy slide images found 97.7% sensitivity and 99.3% specificity, and identified poor-quality scans as an area for improvement. In the reader study behind the FDA authorization, 16 pathologists examined 527 slide images, and the software raised their average sensitivity for cancer on individual slide images by 7.3 percentage points, FDA said. In an earlier reader study of 304 slide images, three pathologists' average sensitivity rose from 74% to 90% with an early version, Paige Prostate Alpha, with no significant change in specificity; in a reader study at one laboratory, four pathologists using the software requested about 20% fewer IHC studies while maintaining accuracy.
- FDA
- FDA granted De Novo authorization DEN200080 on Sept. 21, 2021, and described Paige Prostate as the first AI-based software designed to identify an area of interest on prostate biopsy images. Paige's website calls this FDA approval; FDA's record shows a De Novo grant.
- Limits
- CONFIDENT P was small (82 patients), was not randomized and measured IHC use rather than patient outcomes; Paige funded the Yale validation and the 304-slide reader study, and company employees were among their authors.
Grades: A, randomized evidence of benefit; B, evidence from clinical use; C, accuracy studies only; D, little or no independent evidence. How the grades work. Reviews of the published evidence, not medical advice or an endorsement of any product.