Prostate cancer diagnosis heavily relies on the accurate interpretation of prostate core biopsy whole-slide images. Paige Prostate, a machine-learning algorithm, has emerged as a promising tool to aid pathologists in this crucial task by categorizing images as “suspicious” or “not suspicious” for prostatic adenocarcinoma. This study delves into the evaluation of Paige Prostate within a high-volume academic medical center – Yale Medicine – known for its rigorous diagnostic process involving resident/fellow review, subspecialized genitourinary (GU) pathologist assessment, and consensus conferences for complex cases. The primary objective was to assess Paige Prostate’s effectiveness as both a prescreening tool to streamline workflow by identifying benign cases and as a second-read tool to enhance diagnostic accuracy by detecting potentially missed carcinomas, without specific site-based algorithm adjustments.
Paige Prostate as a Prescreening Tool: Productivity vs. Diagnostic Sensitivity
One potential application of Paige Prostate is to act as a prescreening filter, allowing pathologists to concentrate their expertise on cases flagged as suspicious for malignancy. In our study, utilizing Paige Prostate in this manner would have significantly reduced the workload. Out of 1876 core biopsies, only 589 (31.4%) categorized as suspicious or out-of-distribution would have required manual review. This represents a substantial increase in potential productivity, enabling pathologists to handle a higher volume of cases daily.
However, this efficiency gain comes with a crucial consideration: diagnostic sensitivity. Using Paige Prostate solely as a prescreening tool would have resulted in missing 14 cores with adenocarcinoma. Among these missed malignancies, five were in-focus images, and nine had scan quality issues not flagged by the algorithm as out-of-distribution (OOD). Additionally, six cores with glandular atypia would have been overlooked. Notably, four of the five missed in-focus adenocarcinoma foci were small, measuring 1 mm or less, with two of these also initially missed by manual review. Intriguingly, two of the missed adenocarcinoma foci within the in-focus cores exhibited foamy gland features, a morphological variant known for its deceptively benign cellular characteristics. This observation suggests a potential area for algorithm refinement to improve detection of this specific variant.
Despite these missed cases, it’s important to consider the clinical context. Four of the five missed malignant cores originated from patients who had other cores with larger or higher-grade disease correctly identified by Paige Prostate. In these instances, the missed foci are unlikely to have significantly altered patient management. The fifth missed core was from a patient under surveillance for prior low-grade, low-volume disease, and again, the miss would likely not have changed the clinical course.
Paige Prostate as a Second-Read Tool: Enhancing Diagnostic Accuracy for Subtle Atypia
Beyond prescreening, Paige Prostate’s utility as a second-read tool to validate benign diagnoses was also investigated. In this scenario, the algorithm flagged only 34 slides (1.8% of the total) as suspicious after initial pathologist review deemed them benign. Interestingly, Paige Prostate did not detect any adenocarcinoma foci that were completely missed by the pathologists in the initial review within this study set. However, the algorithm did identify several very small “suspicious” foci.
To further explore these findings, reviewers were asked to re-examine the images flagged as “suspicious” by Paige Prostate, now aware of the algorithm’s concern and with the suspicious areas highlighted. In approximately half of these cases, at least one of the two reviewers changed their initial benign diagnosis to glandular atypia. This shift is clinically significant because glandular atypia, while not definitively malignant, is associated with an increased risk of adenocarcinoma upon re-biopsy. Therefore, Paige Prostate’s ability to highlight such lesions could serve as a valuable adjunct to pathologists, prompting further investigation and potentially leading to earlier detection in subsequent biopsies if warranted.
The “suspicious discrepant” core biopsies identified by Paige Prostate originated from 28 different patients, with a high prevalence of carcinoma (82%) in this group. This prevalence rate is even higher than the overall cancer prevalence (73%) in the entire study cohort, suggesting that Paige Prostate is adept at detecting even subtle atypical glandular changes that may be associated with malignancy. This capability is crucial for identifying Atypical Paige Diagnosis cases that might otherwise be overlooked.
Considerations for Algorithm Improvement and Clinical Integration
Our study, conducted within a highly specialized practice at Yale Pathology, reveals both the potential and areas for improvement of Paige Prostate. While the algorithm may not dramatically increase sensitivity in settings where subspecialized GU pathologists are already involved in multiple reviews, it offers significant promise for enhancing workflow efficiency and potentially improving the detection of subtle atypical lesions.
Several avenues for algorithm enhancement have emerged. Improved identification of out-of-focus scans is crucial to reduce false negative rates arising from technical artifacts. Conversely, refining the algorithm to disregard extremely small, isolated foci (e.g., less than 0.25 mm) could decrease false positive flags. Furthermore, integrating Gleason grading functionality and automated measurement tools for key reporting parameters like biopsy length, cancer length, and percentage involvement would significantly enhance the practical utility of Paige Prostate in routine pathology workflows.
In conclusion, Paige Prostate demonstrates considerable potential as a versatile tool in anatomic pathology practice. Its application as a prescreening tool offers substantial gains in pathologist productivity, while its role as a second-read aid can enhance diagnostic scrutiny, particularly for atypical paige diagnosis and subtle lesions. Further research is warranted to optimize the algorithm, explore its impact on ancillary study utilization, and fully realize its benefits in diverse clinical settings.
