TOPLINE

Researchers developed an AI framework that identified tissue components, classified nine renal cell tumor subtypes, predicted nuclear grade, and stratified survival across multicenter cohorts. The AI-derived risk score outperformed World Health Organization/International Society of Urological Pathology (WHO/ISUP) grading in predicting recurrence-free survival (RFS) and disease-specific survival (DSS).

METHODOLOGY

• A retrospective multicenter study analyzed 11,135 whole-slide images (WSIs) from 7033 patients with renal tumors across 4 medical centers in China and 4 public datasets.
• Researchers used the foundation model Prov-GigaPath to extract histopathologic features and developed a full-stack diagnostic framework employing attention-based multiple instance learning.
• The framework performed 4 sequential tasks: tissue component detection (normal tissue, tumor, necrosis, sarcomatoid differentiation, pseudocapsule), subtype classification to predict nine major renal tumor subtypes, WHO/ISUP patient-level nuclear grading for clear cell renal cell carcinoma (ccRCC) and papillary renal cell carcinoma (pRCC), and prediction of oncologic outcomes using a WSI-based risk score (WRS) from the selected tumor-associated regions.
• Training cohorts comprised 80% of patients from Zhongshan Hospital, Zhangye Hospital, Xiamen Hospital, and the Cancer Imaging Archive datasets, with the remaining 20% forming validation cohorts; Huadong Hospital patients served as an external test cohort.
• Three experienced urologic pathologists re-evaluated all slides according to the 2022 WHO Classification of Tumors of the Urinary System and assigned WHO/ISUP grades and tumor–node–metastasis staging.

TAKEAWAY

• The tissue component detection model achieved area under the receiver operating characteristic curve (AUC) values of 0.990 for normal tissue, 0.982 for tumor tissue, 0.994 for necrosis, 0.967 for sarcomatoid differentiation, and 0.990 for pseudocapsule in the validation cohort, with similar performance in the external test cohort.
• For subtype classification, the model achieved slide-level AUCs ranging from 0.956 to 0.998 across validation cohorts for nine renal tumor subtypes, including 0.990 (validation) and 0.958 (test) for ccRCC, 0.993 (validation) and 0.923 (test) for pRCC, and 0.989 (validation) and 0.995 (test) for chromophobe RCC.
• The nuclear grading model for ccRCC and pRCC achieved an AUC of 0.867 for distinguishing low-grade (WHO/ISUP 1-2) from high-grade (WHO/ISUP 3-4) tumors, with an agreement rate of 86.4% (95% CI, 81.5%-91.0%) with expert consensus, significantly higher than the 81.4% (95% CI, 78.2%-84.5%) agreement rate of general pathologists and subspecialized experts (P < .001).
• The WRS independently predicted overall survival and demonstrated superior prognostic performance compared with WHO/ISUP grading, with concordance indices for RFS of 0.740 (validation) and 0.735 (test) for WRS vs 0.692 (validation) and 0.694 (test) for grading, and for DSS of 0.709 (validation) and 0.743 (test) for WRS vs 0.689 (validation) and 0.721 (test) for grading.

IN PRACTICE

"The proposed framework should be considered a potential decision-support approach rather than a clinically validated autonomous diagnostic system. Prospective studies incorporating real-world deployment and pathologist-AI interaction are required to determine whether it improves diagnostic accuracy, interobserver agreement, reporting efficiency, or pathologist workload," the authors of the study wrote.

SOURCE

The study was led by Ying Xiong, Department of Urology, Zhongshan Hospital, Fudan University, Shanghai, China. It was published online on August 4 in BMC Medicine.

LIMITATIONS

The study did not include all rare renal tumor subtypes due to limited cases for training and validation. Most cases were originally evaluated according to the fourth edition of WHO classification, and some rare subtypes may have been misclassified.

DISCLOSURES

The study was supported by grants from the National Natural Science Foundation of China, International Science and Technology Cooperation Program under the 2023 Shanghai Action Plan for Science, Shanghai Municipal Education Commission Project for Promoting Research Paradigm Reform and Empowering Disciplinary Advancement through Artificial Intelligence, and the Excellent Youth Science Foundation of Fujian Provincial Natural Science Foundation of China. The authors reported no relevant conflicts of interest.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.