Computational Characterization of Lymphocyte Topology on Whole Slide Images of Glomerular Diseases

Key Points Graph topology features of inflammation can enhance prognostication of proteinuric glomerular diseases. Computational image analysis is useful for tissue interrogation and extraction of hidden tissue characteristics. Digital pathology and computer vision allowed for characterization of inflammation patterns beyond human vision. Background The distribution of inflammation in the kidney and its clinical relevance is understudied. This study aimed to computationally quantify lymphocyte topology and test its prediction of disease progression. Methods NEPhrotic syndrome sTUdy NEtwork (NEPTUNE) ( N =333)/Cure Glomerulonephropathy (CureGN) participants ( N =155 focal segmental glomerulosclerosis, N =178 minimal change disease) with available clinical/demographic data and one hematoxylin and eosin–stained whole slide image were included. Cortex and lymphocytes were automatically segmented. A novel graph-based clustering algorithm was applied to identify dense versus sparse lymphocytic habitats, from which 26 pathomic features were extracted to capture cell density, connectivity, clustering, and centrality. The association of these pathomic features with disease progression (40% eGFR decline or KRT) was assessed using ElasticNet-regularized Cox proportional hazards models. Clinical and demographic characteristics and percent of interstitial fibrosis and inflammation were added as potential confounders. Kaplan–Meier survival analysis with log-rank test was used to evaluate risk stratification. Two validation strategies were applied: ( 1 ) training on NEPTUNE with external validation on CureGN data and ( 2 ) using an internal bootstrap validation of the combined datasets for training and validation, respectively. Results Unadjusted analysis: 17 features (65%) retained significance after adjustment for standard clinicodemographic variables, Number of K-core in sparse habitat maintained significance (hazard ratio, 1.29; 95% confidence interval, 1.04 to 1.61) even after adjustment for lymphocyte density, and Average Degree in dense habitat had borderline significance (hazard ratio, 1.25; 95% confidence interval, 1.00 to 1.57) after adjustment for interstitial fibrosis. Multivariable models (clinical/demographic+graph features) achieved validation concordance index of 0.78±0.15 in the CureGN external validation and 0.77±0.06 in the combined internal validation dataset. Time-dependent discrimination showed consistent performance at 3-year (area under the time-varying receiver operating characteristic curve: 0.78 versus 0.76) and 5-year time points (area under the time-varying receiver operating characteristic curve: 0.74 versus 0.76) across validation strategies. Sparse habitat clustering patterns ( Maximum of K-core×Number of K-core in sparse habitat : 88% selection frequency) and dense habitat connectivity ( Average Degree in dense habitat: 47% selection frequency) were consistently identified as robust predictors alongside clinical variables. Conclusions The topologic characterization of lymphocytic inflammation identified immune habitats, capturing the complexity of patterns of inflammation.