Department of Pathology, Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea.
Department of Pathology, Bucheon St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Bucheon, Republic of Korea.
Mod Pathol. 2022 Oct;35(10):1449-1457. doi: 10.1038/s41379-022-01077-9. Epub 2022 Apr 29.
Accurate diagnosis and grading of needle biopsies are crucial for prostate cancer management. A uropathologist-level artificial intelligence (AI) system could help make unbiased decisions and improve pathologists' efficiency. We previously reported an artificial neural network-based, automated, diagnostic software for prostate biopsy, DeepDx Prostate (DeepDx). Using an independent external dataset, we aimed to validate the performance of DeepDx at the levels of prostate cancer diagnosis and grading and evaluate its potential value to the general pathologist. A dataset composed of 593 whole-slide images of prostate biopsies (130 normal and 463 adenocarcinomas) was assembled, including their original pathology reports. The Gleason scores (GSs) and grade groups (GGs) determined by three uropathology experts were considered as the reference standard. A general pathologist conducted user validation by scoring the dataset with and without AI assistance. DeepDx was accurate for prostate cancer detection at a similar level to the original pathology report, whereas it was more concordant than the latter with the reference GGs and GSs (kappa/quadratic-weighted kappa = 0.713/0.922 vs. 0.619/0.873 for GGs and 0.654/0.904 vs. 0.576/0.858 for GSs). Notably, it outperformed the original report, especially in the detection of Gleason patterns 4/5, and achieved excellent agreement in quantifying the Gleason pattern 4. When the general pathologist used AI assistance, the concordance of GG between the user and the reference standard increased (kappa/quadratic-weighted kappa, 0.621/0.876 to 0.741/0.925), while the average slide examination time was substantially decreased (55.7 to 36.8 s/case). Overall, DeepDx was capable of making expert-level diagnosis in prostate core biopsies. In addition, its remarkable performance in detecting high-grade Gleason patterns and enhancing the general pathologist's diagnostic performance supports its potential value in routine practice.
准确的诊断和分级对于前列腺癌的管理至关重要。泌尿科医师级别的人工智能 (AI) 系统可以帮助做出公正的决策并提高病理学家的工作效率。我们之前曾报道过一种基于人工神经网络的自动化前列腺活检诊断软件，即 DeepDx Prostate (DeepDx)。我们使用一个独立的外部数据集，旨在验证 DeepDx 在前列腺癌诊断和分级水平上的性能，并评估其对普通病理学家的潜在价值。该数据集由 593 张前列腺活检的全切片图像组成（130 张正常和 463 张腺癌），并包含其原始病理报告。由三位泌尿科病理学家确定的 Gleason 评分 (GS) 和分级组 (GG) 被视为参考标准。一名普通病理学家在有和没有人工智能辅助的情况下对数据集进行评分，以进行用户验证。DeepDx 在前列腺癌检测方面的准确性与原始病理报告相似，但与参考 GG 和 GS 的一致性更高（GG 的kappa/二次加权 kappa 值分别为 0.713/0.922 和 0.619/0.873，GS 分别为 0.654/0.904 和 0.576/0.858）。值得注意的是，它的表现优于原始报告，特别是在检测 Gleason 模式 4/5 方面，并在定量评估 Gleason 模式 4 方面达成了极好的一致性。当普通病理学家使用人工智能辅助时，用户和参考标准之间 GG 的一致性增加（kappa/二次加权 kappa 值分别为 0.621/0.876 和 0.741/0.925），而平均切片检查时间则大大减少（从 55.7 秒/例减少至 36.8 秒/例）。总的来说，DeepDx 能够对前列腺核心活检进行专家级别的诊断。此外，它在检测高级别 Gleason 模式方面的出色表现以及提高普通病理学家的诊断性能，支持其在常规实践中的潜在价值。
