To obtain a more complete understanding of material microstructure at the nanoscale and to gain profound insights into their properties, there is a growing need for more efficient and precise methods that can streamline the process of 3D imaging using a transmission electron microscope (TEM) and reconstruction, thereby minimize the need for manual intervention. This interdisciplinary project focuses on the development of an automatic stereo-vision reconstruction method that utilizes deep-learning neural networks. The aim is to obtain the 3D distribution of significant one-dimensional curvilinear defects, specifically dislocations, by leveraging just two scanning transmission electron microscope (STEM) images captured from different viewing angles. As part of this project, the imaging conditions were optimized with the specific objective of ensuring that the captured images are optimal for the subsequent reconstruction process. The optimization aimed to maximize the quality and usefulness of the images for the reconstruction of the 3D distribution of dislocations. The specific characteristics of these images were considered, and a quantitative assessment was conducted using topological data analysis (TDA) on the neural network outputs. The results demonstrated a significant reduction in imaging and post-processing times compared to conventional techniques such as electron tomography. The study revealed the robustness of neural networks in handling noise levels in images, highlighting their ability to extract relevant information from noisy data. Further optimization of the "tilt-less" technique has resulted in an enhanced stereo-angle for imaging and improved reconstruction quality. By incorporating a pixelated detector, it was possible to achieve a stereo-angle of 9.32 degrees, while also gaining better control over the tilt angle. These advancements have not only been applied successfully to reconstruct the 3D distribution of dislocations but also to assess the 3D distribution of nanoparticles under cryogenic conditions. As a result, this approach has significantly reduced the acquisition time, as only a single acquisition is now required compared to the previous method of using tomography with multiple images. The developed techniques in this work enable real-time observation of dynamic processes within the 3D volume of materials, opening new frontiers of exploration.