A multi-structure 3D multiphoton liver microscopy dataset integrating real, physics-based, and GAN-simulated volumes for benchmarking bioimage analysis

Three-dimensional (3D) microscopy enables quantitative analysis of tissue architecture; however, rigorous development, validation, and comparison of volumetric image-analysis methods depend on publicly available benchmark datasets with validated ground truth, which remain limited due to the technical challenges of 3D annotation. Here, we present a multi-structure 3D liver microscopy dataset designed for systematic benchmarking of bioimage analysis methods. The resource comprises 44 volumetric image stacks derived from multiphoton microscopy, together with manually curated segmentation masks, idealized isotropic binary tissue models, and two complementary simulated image sets generated using physics-based image formation modeling and a 3D CycleGAN framework. The complete dataset, including raw and processed images, segmentation masks, experimental point spread functions, simulation pipelines, trained models, and prediction outputs, amounts to approximately 500 GB of publicly available data. The volumes capture four principal hepatic structures spanning distinct spatial scales and morphologies: cell borders, tubular structures (bile canaliculi and sinusoids), and nuclei. Controlled signal-to-noise ratios, depth-dependent intensity variations, and experimentally measured point spread functions are incorporated to reproduce realistic imaging conditions. By integrating real acquisitions, analytically defined ground truth, and simulated volumes with controlled degradations, this resource enables reproducible evaluation of isotropic reconstruction, multi-structure segmentation, instance detection, and image restoration methods in 3D microscopy. All data, trained models, and processing workflows are openly available to support transparent benchmarking and methodological development in volumetric bioimage analysis.