MONKEY Challenge: Machine-learning for Optimal detection of iNflammatory cells in KidnEY transplant biopsies
文献摘要
Purpose Accurate quantification of inflammatory cells in Periodic Acid-Schiff (PAS)-stained kidney transplant biopsies is essential for Banff classification but remains subjective and labor-intensive. To address this, the MONKEY challenge benchmarked deep learning (DL) methods for automated detection of mononuclear leukocytes and their subtypes. Materials and methods A multicenter dataset of 118 biopsies with co-registered PAS and immunohistochemically restained slides, comprising over 131,000 annotated cells, was used. The algorithms were evaluated on hold-out validation and test set using Free-response Receiver Operating Characteristic (FROC) analysis, alongside precision, recall, and F1-scores. Results Across diverse architectures, performance converged, with the top models achieving F1-scores up to 0.75 for overall detection. Lymphocyte detection performed well (F1 up to 0.78), whereas monocyte detection remained more challenging (F1 up to 0.47). To explore the potential clinical applicability of automated inflammatory cell detection, the winning model was integrated with renal tissue segmentation models, enabling compartment-specific evaluation and quantitative assessment of Banff-related lesions. In an independent multicenter cohort of 180 kidney transplant biopsies, DL-derived quantitative measurements demonstrated strong correlations with consensus pathologist scoring for several inflammation-related Banff lesions, particularly total inflammation and interstitial inflammation. Conclusions This study establishes the first multicenter, PAS-based, IHC-validated benchmark for inflammatory cell detection and provides a clinically grounded resource to support the development of robust, generalizable AI systems for transplant pathology. Beyond benchmark performance, our extended tissue-specific analyses demonstrate that inflammatory cell detection on routine PAS slides can be anatomically contextualized and translated into quantitative measurements that align with conventional Banff lesion assessment, providing a foundation for future AI-assisted evaluation of kidney transplant biopsies.