Many Cryo-EM datasets contain structural heterogeneity due to functional or nonfunctional dynamics that conventional reconstruction methods may fail to resolve. Here we propose a new method, OPUS-DSD (Deep Structural Disentanglement), which can reliably reconstruct the structural landscape of cryo-EM data by directly translating the 2D cryo-EM images into 3D structures. The method adopts a convolutional neural network and is regularized by a latent space prior that encourages the encoding of structural information. The performance of OPUS-DSD was systematically compared to a previously reported method, cryoDRGN, on synthetic and real cryo-EM data. It consistently outperformed existing methods, resolved large or small structural heterogeneity, and improved the final reconstructions of tested systems even on highly noisy cryo-EM data. The results have shown that OPUS-DSD should be particularly suitable for cases in which the high structural flexibilities cannot easily be represented by rigid-body movements. Therefore, OPUS-DSD represents a valuable tool that can not only recover functionally-important structural dynamics missed in a traditional cryo-EM refinement, but also improve the final reconstruction by increasing homogeneity in a dataset.