Probabilistic failure prediction of composite materials considering its microstructure randomness is of critical importance for Integrated Computational Material Engineering (IMCE). The dual high-dimensionality model for both physics-modeling and uncertainty quantification is required to accurately represent the voxel-level behavior, such as fracture initiation. This dual high-dimensionality requirement causes a great challenge for computational efficiency, especially for probabilistic predictions. A novel framework is proposed in this paper to address this extreme high dimensional problem by integrating physics-modeling with the recent advancement of deep learning techniques for imaging analysis. Thus, the material fracture prediction problem is treated as a hybrid physics-constrained imaging regeneration problem. The physics-based deep learning method can serve as a surrogate model for probabilistic analysis and super computational efficiency is observed. The proposed methodology includes three major parts. First, an explicit analytical random microstructure quantification model is proposed using a non-Gaussian random field expansion technique. The major advantage is that the random microstructure can be rigorously quantified and efficiently reconstructed for future learning. Then, the reconstructed microstructure can be combined into the lattice particle model (LPM) to predict the failure of the materials, which has the benefits that spatial discontinuities can be avoided compared with the conventional continuum-based computation models. The LPM can be time-consuming, especially for large-scale problems. Next, a hybrid physics-based deep learning method is trained to serve as the surrogate of the mechanics model. The time-consuming nonlinear crack initiation and propagation analysis are replaced with a set of imaging training problems. The feasibility and efficiency of the framework are investigated by several probabilistic failure analysis demonstrated examples. Finally, several potential future research works are discussed following the conclusions.