#6541 GLOMERULONEPHRITIS DIAGNOSIS BY MACHINE LEARNING ON PERIODIC ACID-SCHIFF (PAS) WHOLE SLIDE IMAGES

Abstract Background and Aims Machine learning (ML) holds great promise for improving diagnostics, prognostication theranostics in nephropathology. So far, applications have not gone much further than segmentation of tissue compartments on whole slide images (WSIs) paraffin sections. As a proof-of-concept study, we describe the development diagnostic classifier glomerulephritis based expert-annotated or automatically segmented glomerular transections from periodic-acid Schiff (PAS) sections only. Method A total n = 350 biopsies 5 institutions with 12 classes glomerulonephritis IgA nephropathy (IgAN), membranous (Membranous), anti-glomerular basement membrane antibody GN (ABMGN), infection-associated (IAGN), ANCA-associated (ANCA-GN), idiopathic membranoproliferative (MPGN), SLE class IV (SLE-GN-IV), cryglobulinemic (CryoGN), C3 (C3-GN), dense deposit disease (DDD), fibrillary (FibrillaryGN) proliferative monoclonal immunoglobulin deposits (PGNMID) were included study their respective PAS Glomerular by nephropathologist our own transformer-based model trained 100 thrombotic microangiopathies range vascular, vasculitic diseases closely resembling/mimicking microangiopathies. For classification, divided cohort into folds internal cross-validation, performed sample size augmentation various methods (including shifts resolution/scale, AutoAugment others) proprietary self-attention-based MILx architecture an EfficientNet backbone selection crop batches soft Markov chain Monte Carlo sampling semi-supervised fashion, labels each biopsy. We compared performance both crops recently published benchmark (CLAM) multiple-instance histopathology. Results Automatic was excellent mean AUC sensitivity (mean average recall) over all at 0.904, near perfect specificity (0.994), as expected best Membranous, worst ABMGN. Classification inputs had balanced accuracy 0.84, metrics descending order 0.97 0.89 ABMGN, 0.88 IgAN, 0.86 Fibrillary, 0.83 MPGN, 0.80 ANCA-GN, 0.79 DDD, 0.78 PGNMID, 0.75 IAGN, 0.73 SLE-GN-IV CryoGN, 0.67 C3-GN. Performance similar input. On this dataset, outperformed CLAM entire WSIs well 0.72) significant margin. Conclusion This proof-of-concept-study indicates that nephropathology-specific architectures like can be complex tasks relatively small biopsy cohorts. should able to deliver end-to-end-pipeline other training sets case-labels provided trusted only minimal expert labeling annotation required. PAC HQ contributed equally work.