DUCK-Net: Automated deep learning segmentation of Ductular Reactions in murine liver injury captures multicellular niche dynamics from H&E morphology

Abstract Ductular Reactions (DRs) are dynamic and complex multicellular responses that occur as a result of various hepatic injuries. Precise identification and quantification of the extent of DRs is a cornerstone of pre-clinical modelling of liver disease, with links to inflammation, fibrosis, regeneration, and disease severity. Here, we apply a deep learning model, Deep Understanding Convolutional Kernel or ‘DUCK-Net’, to the automated detection and segmentation of DRs in whole-slide histopathological images of murine models of liver damage. Following annotation of a training dataset by a specialist liver histopathologist, we demonstrate accelerated performance and accurate detection, achieving a mean Dice coefficient (model-expert segmentation overlap) of 85.4% and a specificity of 98%, indicating minimal false positives. Evaluation of model validity and utility was achieved with a histological time course of cholestatic injury and recovery using 3,5-Diethoxycarbonyl-1,4-Dihydrocollidine diet (DDC) in mice. When assessed against a multiple linear regression model incorporating core epithelial and stromal components of the DR as quantified using IHC, DUCK-Net predicted the spatiotemporal response to injury and repair/resolution with a coefficient of determination ( R 2 ) of 0.88. Moreover, DUCK-Net kinetics strongly correlated with published spatial transcriptomic (Stereo-seq) analysis of the DDC model, demonstrating that H&E-based segmentation captures molecular DR dynamics comparable to, or exceeding that of individual IHC markers without the need for immunostaining. DUCK-Net provides a novel and accessible platform for rapid, accurate histological quantification of liver injury reflective of the matrix-rich, multicellular regenerative niche observed in DRs.