Next-Generation Morphometry for pathomics-data mining in histopathology

Abstract

Pathology diagnostics relies on the assessment of morphology by trained experts, which remains subjective and qualitative. Here we developed a framework for large-scale histomorphometry (FLASH) performing deep learning-based semantic segmentation and subsequent large-scale extraction of interpretable, quantitative, morphometric features in non-tumour kidney histology. We use two internal and three external, multi-centre cohorts to analyse over 1000 kidney biopsies and nephrectomies. By associating morphometric features with clinical parameters, we confirm previous concepts and reveal unexpected relations. We show that the extracted features are independent predictors of long-term clinical outcomes in IgA-nephropathy. We introduce single-structure morphometric analysis by applying techniques from single-cell transcriptomics, identifying distinct glomerular populations and morphometric phenotypes along a trajectory of disease progression. Our study provides a concept for Next-generation Morphometry (NGM), enabling comprehensive quantitative pathology data mining, i.e., pathomics.

Fig. 1. Flowchart of the patient cohorts and the integration of our framework for large-scale histomorphometry (FLASH) into a digital pathology workflow.

a Overview of the cohort refinement process. Cases and whole-slide images (WSIs) were excluded based on predefined criteria on case- and slide-level. 1043 cases from five cohorts and 1743 WSIs were included in this study. b Integration of FLASH into the digital pathology workspace. FLASH combines deep learning-based segmentation with bioinformatics analysis of quantitative morphometric features. The framework consists of two convolutional neural networks (CNNs) for tissue and structure segmentation, computational feature extraction and Next-Generation Morphometry (NGM) analysis. IgAN IgA nephropathy, WSIs whole-slide images, PAS periodic acid schiff.

Fig. 2. NGM-derived glomerular features reveal distinct morphometric patterns in native kidney diseases and different clinical conditions, such as nephrotic range proteinuria and reduced kidney function.

a–a″″ Segmentation visualisations of glomeruli in major glomerular injury patterns (images stem from the internal AC_B cohort excluding training samples). b Visual representation for calculation of glomerular tuft circularity as an example of one of the extracted morphometric features. c Comparison of glomerular tuft area [μm²] on instance-level with 11,077 instances in different native kidney diseases from the AC_B cohort. Glomeruli from biopsies without pathological findings were used as a control (depicted in grey). d Feature analysis of glomerular tuft area on instance-level based on nephrotic range proteinuria in all native biopsies from the AC_B cohort, d′ for glomeruli from biopsies diagnosed with minimal change disease (MCD) or membranous glomerulonephritis (GN) and d″ for glomeruli with large proteinuria from the external KPMP cohort. Visualisations highlight the increase in glomerular tuft area in cases with nephrotic range proteinuria. e Comparison of glomerular tuft circularity on instance-level between cases of MCD and membranous GN with or without nephrotic range proteinuria. f Analysis of glomerular tuft shape on instance-level based on reported estimated glomerular filtration rate in all native biopsies from our internal biopsy cohort including additional visualisation examples. Scale bar size is 100 µm. Source data are provided as a Source Data file. GN glomerulonephritis, Seg. segmental, HTN hypertensive nephropathy, DN diabetic nephropathy, IgAN IgA nephropathy, MCD minimal change disease, lupus lupus nephritis, membranous membranous glomerulonephritis, Pauci Pauci-immune glomerulonephritis, eGFR estimated glomerular filtration rate.

Fig. 3. NGM-derived features of tubules and arteries are associated with pathologist derived scoring.

a–a″″ Segmentation visualisations of tubules with large variation in size and shape in various diseases and morphological injury patterns present in the patch. Visualisations stem from the internal AC_B cohort excluding training samples. b Visual representation of feature calculation of tubular diameter and tubular distance. c Feature analysis of tubular diameter on instance-level based on the quantified amount of interstitial fibrosis and tubular atrophy (IFTA) of all biopsies with reported IFTA from the AC_B cohort. c′ Analysis of tubular diameter on instance-level based on the measured estimated glomerular filtration rate (eGFR) of all native biopsies from our internal biopsy cohort. d Feature analysis of tubular distance summarised on patient-level based on the quantified IFTA of all biopsies with reported IFTA from the internal biopsy cohort. d′ Analysis of tubular distance summarised on patient-level based on the measured eGFR of all native biopsies from our internal biopsy cohort. Scale bar size is 100 µm. Source data are provided as a Source Data file. DN diabetic nephropathy, ABMR antibody-mediated rejection, d_max maximum instance diameter, dist_min minimum instance distance, IFTA interstitial fibrosis and tubular atrophy, eGFR estimated glomerular filtration rate.

Fig. 4. NGM-derived quantitative features are predictive of disease progression in IgA nephropathy (IgAN).

a Comparison of five predictive digital biomarkers summarised at patient-level based on reaching the defined composite endpoint, i.e., end-stage kidney disease and/or halving of initial estimated glomerular filtration rate (eGFR) within 15 years after biopsy. b Univariate Cox proportional hazards models for 644 patients of the five predictive features summarised at patient-level including 95% confidence intervals. In three cases no glomerular tuft was segmented, and no shape features were calculated. Cumulative events for each group in the univariate Cox proportional hazard models are provided in Supp. Table 9. c Hazard ratios (centre) and their 95% confidence interval (error bars) from the univariate Cox proportional hazard models of the respective features. Source data are provided as a Source Data file. ESKD end-stage kidney disease, eGFR estimated glomerular filtration rate, HR hazard ratio, CI confidence interval.

Fig. 5. Pseudotime analysis of NGM-derived glomerular features identifies distinct glomerular groups along a trajectory of disease progression in IgA nephropathy (IgAN).

a–a″ Diffusion map embedding of 24,227 glomerular instances with 14 morphometric features with IgAN based on the reported estimated glomerular filtration rate (eGFR) [ml/min/1.73 m²]. b Diffusion map of glomerular instances with pseudotime indicating ordering of glomerular instances along their progression from healthy to diseased. b′ Visualisation of glomerular phenotypes along the pseudotime. c Scaled feature expression heatmap including eGFR along the pseudotime trajectory. d Morphometric progression of glomerular instances in clinical subgroups based on the overall reported eGFR. Scale bar size is 100 µm. Source data are provided as a Source Data file. eGFR estimated glomerular filtration rate, Dim diffusion map, IgAN IgA nephropathy.