Deep learning-driven immunohistochemical analysis of renal lymphatics for chronic kidney disease: bioinformatic and...

Renal lymphatic vessel density is clinically relevant in kidney disease but is still assessed by slow, subjective visual estimation. We evaluated a weakly supervised, attention-based multiple-instance learning framework for automated detection and quantification of renal lymphatic vessel density on D2-40-stained whole-slide images (WSIs).

Two independent internal datasets from Tongji Hospital were collected, including 198 cases of chronic kidney disease (CKD) and 50 cases of hypertensive nephropathy (HTN). All biopsies were immunohistochemically stained for D2-40 and digitized as WSIs. Pathologists provided only slide-level labels (D2-40 high vs. D2-40 low). Tissue regions were automatically segmented, tiled into patches, and encoded using a pretrained convolutional neural network. Patch embeddings were then analyzed with a clustering-constrained attention multiple-instance learning (CLAM) model. Unlike conventional multiple-instance learning (MIL) methods that only weight instances, CLAM jointly performs attention-based instance selection and instance-level clustering to distinguish positive from negative evidence within each slide, yielding more discriminative slide-level features and interpretable attention maps. Performance was compared with a classic MIL model trained on the same features.

CLAM achieved area under the receiver operating characteristic curves of 0.942 and 0.858 on the CKD and HTN datasets, respectively, outperforming classic MIL (0.866 and 0.801). Attention maps highlighted lymphatic-rich regions consistent with renal pathologists' assessments.

This clustering-constrained, attention-based weakly supervised framework enables fully automated, reproducible quantification of renal lymphatic vessel density from WSIs, providing renal pathologists with rapid visual and numerical support for diagnosis and risk stratification in CKD and HTN.