An artificial intelligence model for nuclear grading of clear cell renal cell carcinoma using whole slide images: a retrospective, multicentre, diagnostic study.

Abstract

Background: The pathological assessment of International Society of Urological Pathology (ISUP) nuclear grading is crucial for the management of clear cell renal cell carcinoma (ccRCC). We aimed to develop an artificial intelligence (AI)-based, high-efficiency, and high-accuracy ccRCC ISUP Grading Diagnostic System (RIGDAS) and evaluate its clinical application value.

Methods: In this multicentre, retrospective, diagnostic study, consecutive ccRCC patients who underwent partial or complete nephrectomy between 1 June 2014 and 1 June 2024 across three Chinese hospitals and two public cohorts were included. Pathological slides from these surgeries were collected and digitized into whole slide images for model development and validation. The primary endpoint was the area under the receiver operating characteristic curve (AUC) of RIGDAS. Additionally, the performance and review time of pathologists assisted with RIGDAS were evaluated.

Results: A total of 5,697 slides from 1,807 ccRCC patients were collected and digitized for training and validating RIGDAS. Across the training and validation datasets, RIGDAS achieved an AUC ranging from 0.943 (95% confidence interval [CI], 0.927-0.971) to 0.980 (0.960-1.989). In the human-AI comparison and collaboration study, RIGDAS achieved an accuracy (0.930 [0.907-0.951]) that was 3.3-4.3% higher than the accuracy of two junior pathologists (0.897 [0.883-0.916], P = 0.004; 0.887 [0.871-0.904], P = 0.001) and was comparable to the accuracy of two senior pathologists (0.960 [0.948-0.977] and 0.970 [0.961-0.986], both P > 0.05). Furthermore, RIGDAS significantly improved the diagnostic accuracy of the two junior pathologists to the level of the senior pathologists (P > 0.05) and greatly reduced the slide review time for all four pathologists (20.5-45.1%, all P < 0.0001).

Conclusion: RIGDAS demonstrated decent ability in diagnosing ISUP nuclear grading in ccRCC, reducing the likelihood of misdiagnosis by pathologists, and decreasing the time required for pathological slide review, highlighting its potential for clinical application.