This thesis presents a suite of deep learning frameworks designed to address key challenges in the analysis of Electron Microscopy (EM) images of cancer cells, which are vital for understanding cancer progression and therapy resistance. The research focuses on three primary domains: (1) Image Quality Assessment and Denoising, where automated metrics and deep learning models validate faster, safer EM sample preparation protocols without compromising image quality; (2) Semi-supervised Semantic Segmentation, leveraging limited manual annotations to train models for accurate 3D segmentation of nuclei and nucleoli, essential for cancer diagnostics; and (3) Object Detection and Unsupervised Segmentation, integrating advanced detection and segmentation methods to analyze complex organelles such as mitochondria and endosomes at nanoscale resolution. The findings demonstrate the efficacy of these approaches in reducing annotation bottlenecks and enhancing the robustness of EM image analysis. By automating key aspects of image processing, this work contributes significantly to accelerating cancer research and supporting clinical applications.