Perineural invasion

Algorithm

Prostate cancer

Computer science

Receiver operating characteristic

Radiology

Internal medicine

Artificial intelligence

Prostate

Medicine

Gold standard (test)

Cancer

Pathology

The Lancet Digital Health

<h2>Summary</h2><h3>Background</h3> There is high demand to develop computer-assisted diagnostic tools to evaluate prostate core needle biopsies (CNBs), but little clinical validation and a lack of clinical deployment of such tools. We report here on a blinded clinical validation study and deployment of an artificial intelligence (AI)-based algorithm in a pathology laboratory for routine clinical use to aid prostate diagnosis. <h3>Methods</h3> An AI-based algorithm was developed using haematoxylin and eosin (H&E)-stained slides of prostate CNBs digitised with a Philips scanner, which were divided into training (1 357 480 image patches from 549 H&E-stained slides) and internal test (2501 H&E-stained slides) datasets. The algorithm provided slide-level scores for probability of cancer, Gleason score 7–10 (<i>vs</i> Gleason score 6 or atypical small acinar proliferation [ASAP]), Gleason pattern 5, and perineural invasion and calculation of cancer percentage present in CNB material. The algorithm was subsequently validated on an external dataset of 100 consecutive cases (1627 H&E-stained slides) digitised on an Aperio AT2 scanner. In addition, the AI tool was implemented in a pathology laboratory within routine clinical workflow as a second read system to review all prostate CNBs. Algorithm performance was assessed with area under the receiver operating characteristic curve (AUC), specificity, and sensitivity, as well as Pearson's correlation coefficient (Pearson's <i>r</i>) for cancer percentage. <h3>Findings</h3> The algorithm achieved an AUC of 0·997 (95% CI 0·995 to 0·998) for cancer detection in the internal test set and 0·991 (0·979 to 1·00) in the external validation set. The AUC for distinguishing between a low-grade (Gleason score 6 or ASAP) and high-grade (Gleason score 7–10) cancer diagnosis was 0·941 (0·905 to 0·977) and the AUC for detecting Gleason pattern 5 was 0·971 (0·943 to 0·998) in the external validation set. Cancer percentage calculated by pathologists and the algorithm showed good agreement (<i>r</i>=0·882, 95% CI 0·834 to 0·915; p<0·0001) with a mean bias of −4·14% (−6·36 to −1·91). The algorithm achieved an AUC of 0·957 (0·930 to 0·985) for perineural invasion. In routine practice, the algorithm was used to assess 11 429 H&E-stained slides pertaining to 941 cases leading to 90 Gleason score 7–10 alerts and 560 cancer alerts. 51 (9%) cancer alerts led to additional cuts or stains being ordered, two (4%) of which led to a third opinion request. We report on the first case of missed cancer that was detected by the algorithm. <h3>Interpretation</h3> This study reports the successful development, external clinical validation, and deployment in clinical practice of an AI-based algorithm to accurately detect, grade, and evaluate clinically relevant findings in digitised slides of prostate CNBs. <h3>Funding</h3> Ibex Medical Analytics.

An artificial intelligence algorithm for prostate cancer diagnosis in whole slide images of core needle biopsies: a blinded clinical validation and deployment study