Abstract
INTRODUCTION: Histological analysis of kidney biopsies is crucial in diagnosing kidney diseases and predicting clinical outcomes. Recently developed pathology foundation models, pretrained on large-scale pathology datasets, have demonstrated excellent performance in various downstream applications. This study evaluated the utility of pathology foundation models combined with multiple instance learning (MIL) for kidney pathology analysis. METHODS: We used 242 hematoxylin and eosin (H&E)-stained whole slide images (WSIs) from the Kidney Precision Medicine Project (KPMP) and Japan-Pathology Artificial Intelligence Diagnostics Project (JP-AID) databases as the development cohort, comprising 47 healthy controls, 35 acute interstitial nephritis, and 160 diabetic kidney disease (DKD) slides. External validation was performed using 83 WSIs from the University of Tokyo Hospital (UT dataset). Diagnoses were based on adjudicated diagnoses (KPMP) or expert pathologists-derived diagnoses (JP-AID and UT). Pretrained pathology foundation models were utilized as patch encoders and compared with ImageNet-pretrained ResNet50. RESULTS: In internal validation, all foundation models outperformed ResNet50, achieving area under the receiver operating characteristic curve (AUROC) over 0.980. In external validation, the performance of ResNet50 markedly dropped (AUROC = 0.768), whereas all foundation models maintained higher performance (AUROC over 0.800). Visualization with attention heatmaps confirmed that foundation models accurately recognized diagnostically relevant structures. Additionally, foundation models outperformed ResNet50 in predicting severe proteinuria among DKD cases from KPMP dataset. CONCLUSION: We successfully integrated pathology foundation models with MIL to achieve robust diagnostic performance, even when trained on a relatively small dataset, highlighting their potential for real-world clinical applications. Key words: artificial intelligence, renal pathology, foundation model, multiple instance learning.