A fundamental understanding of the microstructure–property relationships (MPRs) is crucial for optimising the performances and functionality of multiphase composites. Image-based poro/micro-mechanical modelling offers a powerful non-invasive method to explore MPRs, but the inherent randomness in multiphase composites often necessitates extensive datasets of 3D digital microstructures for reliable statistical analysis. This paper presents a cost-effective machine learning-based framework to efficiently reconstruct numerous virtual 3D microstructures from a limited number of 2D real exemplars, bypassing the prohibitive costs associated with volumetric microscopy for opaque composites. This innovative framework leverages feedforward neural networks to encode morphological statistics in 2D exemplars, referred to as the statistics-encoded neural network (SENN), providing an accurate statistical characterisation of complex multiphase microstructures. Utilising the SENN-based characterisation, 3D morphological statistics can be inferred from 2D measurements through a 2D-to-3D morphology integration scheme, and then statistically equivalent 3D microstructures are synthesised via Gibbs sampling. This framework further incorporates hierarchical characterisation and multi-level reconstruction approaches, allowing for the seamless capture of local, regional, and global microstructural features across multiple length scales. Validation studies are conducted on three representative multiphase composites, and morphological similarity between the reconstructed and reference 3D microstructures is evaluated by comparing a series of morphological descriptors. Additionally, image-based meshing and pore/micro-scale simulations are performed on these digital microstructures to compute effective macroscopic properties, including stiffness, permeability, effective diffusivity, and thermal conductivity tensors. Results reveal strong statistical equivalence between the reconstructed and reference 3D microstructures in both morphology and physical properties, confirming the SENN-based framework is a high-fidelity tool to reconstruct multiphase microstructures for image-based poro/micro-mechanical analysis.