深度学习量化1型糖尿病自然史中的内分泌细胞与CD3+T细胞

原文

AIMS/HYPOTHESIS: Histopathological analysis in type 1 diabetes presents challenges in achieving precise characterisation with cellular quantification from whole-slide images (WSIs). To ensure a comprehensive understanding of changes in endocrine cells, this study leveraged a pre-trained deep learning-assisted analysis workflow to enhance the understanding of histopathological features of type 1 diabetes development across pancreases from control, autoantibody-positive and type 1 diabetes organ donors. METHODS: Three pancreatic sections (head, body and tail regions), stained for insulin (INS), glucagon (GCG) and CD3 by immunohistochemistry, were analysed from 32 autoantibody-negative control donors, 12 single-autoantibody-positive (sAAb+) donors, eight multi-autoantibody-positive (mAAb+) donors, six donors with recent-onset type 1 diabetes (0-1 years disease duration) and 19 donors with longstanding type 1 diabetes. Endocrine cell groups (i.e. clusters [<1000μm2] and islets [≥1000μm2]) were segmented by a pre-trained Segment Anything Model, followed by precise segmentation for INS+ and GCG+ cell regions within these structures using a QuPath pixel classifier. CD3+ cells located within a 20 µm periphery of each endocrine cell group were quantified. Ordinal regression was applied to assess disease stage-associated patterns in quantified predictors. The Kruskal-Wallis test was used to compare across the five donor groups. For pairwise comparisons, Wilcoxon rank-sum tests with Bonferroni correction were conducted. RESULTS: From a total of 231 WSIs from 77 donors, >82,000 islets and >26,000 clusters were analysed. In ordinal regression, fractional INS and GCG areas were the most significant predictors of type 1 diabetes progression. CD3+ immune cell infiltration in islets demonstrated a high association with type 1 diabetes progression. Infiltration in both islets and clusters peaked at disease onset before declining, suggesting that these structures are synchronised targets within the autoimmune process. Insulitic clusters were evident even prior to the onset of type 1 diabetes, underscoring the early involvement of these structures in the autoimmune process. CONCLUSIONS/INTERPRETATION: The deep learning-powered approach enabled our study to include clusters of endocrine cells scattered throughout WSIs, providing precise quantitative evidence of cluster-level infiltration. The identification of autoimmune patterns in both islets and clusters, alongside the quantification of beta and alpha cells across donor groups and pancreatic regions, offers a more detailed understanding of type 1 diabetes pathogenesis. Our findings provide robust evidence of cluster-level infiltration even before type 1 diabetes onset, supporting early intervention efforts to preserve beta cells.