Deep learning enables feature extraction of 3D collagen architecture in cleared fibrotic tissues

ABSTRACT Light-sheet fluorescence microscopy enables deep optical sectioning of large, cleared biological tissues. However, effective clearing of collagen-rich tissues remains a persistent technical challenge. Moreover, standardized workflows integrating three-dimensional imaging with computational analysis of collagen architecture are currently unavailable. Here, we present an integrated pipeline combining optimized tissue clearing, volumetric light-sheet imaging, and deep learning-based feature extraction of collagen architecture. Using experimental desmoid tumors as a proxy for collagen-dense tissues, we optimised DISCO-based clearing incorporating Fast Green FCF for collagen labelling. We achieved optical transparency and full 3D visualization of collagen architectures in desmoid tumors, human skin biopsies, and fibrotic mouse lung & liver tissues, including FFPE samples. Using the Benchtop mesoSPIM platform, we acquired high-resolution volumetric datasets and validated multimodal collagen imaging through two-photon microscopy with the Schmidt-Voigt objective. To enable automated feature extraction from these large volumetric datasets, we developed ColNet, a U-Net model for automated collagen fiber segmentation. ColNet demonstrated robust generalization across diverse human and mouse tissues without retraining or hyperparameter adjustment. This integrated workflow provides a foundation for future quantitative assessment of cell-extracellular matrix dynamics in a fibrotic context.