<p>COVID-19 created an unusually large and rapidly evolving body of chest imaging research, making it a useful case study for examining how optimized deep learning (DL) systems behave under urgent clinical demand, limited labels, heterogeneous datasets, and deployment pressure. Accordingly, the review focuses primarily on DL-based COVID-19 detection and assessment from chest X-ray (CXR) and computed tomography (CT), while using related chest diseases such as pneumonia, tuberculosis, lung cancer, and chronic obstructive pulmonary disease (COPD) as contextual comparators rather than as equally weighted disease-specific review categories. The review synthesizes convolutional neural networks, transfer learning, ensemble and hybrid designs, transformer-based architectures, generative approaches for data scarcity, and optimization strategies such as hyperparameter tuning, evolutionary algorithms, and data augmentation. It further examines public CXR and CT datasets, preprocessing pipelines, evaluation protocols, explainability methods, and privacy-preserving learning schemes. The proposed taxonomy relates modality, architecture, optimization strategy, and validation design to reported performance. The evidence indicates that optimized DL pipelines can achieve high internal performance in COVID-19 imaging tasks; however, reported gains are strongly conditioned by dataset size, class balance, single- versus multi-center acquisition, modality, and validation protocol. Persistent barriers include shortcut learning, limited external validation, uneven reporting of preprocessing, insufficient explanation fidelity, and regulatory requirements for clinically deployed AI systems. The review concludes by identifying concrete priorities for robust, interpretable, privacy-aware, and clinically transferable COVID-19 chest imaging AI.</p>