Over the past several years, new advances in computing hardware and artificial intelligence techniques have allowed deep learning to rapidly develop as a key tool in a broad range of fields. In medical imaging, significant attention has been devoted to exploring how these technologies can improve radiological workflow, including more efficient and more accurate image reading and serving as a rapid, objective reader acting concurrently with human radiologists. However, several challenges exist in applying typical deep learning technologies to CT scans. In this dissertation research, we consider three thoracic CT use cases and evaluate novel deep learning techniques to improve clinical utility. The first aim of this dissertation was to develop a deep learning algorithm to evaluate coronary artery calcification (CAC) on low-dose thoracic CT (LDCT) scans. Coronary heart disease is the leading cause of death globally and CAC scores serve as a strong predictor for adverse events related to coronary heart disease. To automatically score LDCT scans, we developed a novel image segmentation network, CACU-Net, which identifies CAC on LDCT scans and classifies lesions based on the coronary artery branch. CACU-Net was able to identify which LDCT scans and individual arteries contain CAC and classify scans into clinically relevant categories based on severity of CAC, outperforming similar segmentation approaches. A second algorithm was developed to detect emphysema on LDCT scans using a transfer attention-based multiple instance learning (TAMIL) approach. This novel technique evaluates slices individually using a transfer learning feature extraction algorithm that requires no additional network training. The slice features are then aggregated through a learned attention-based pooling method that both improves performance and provides interpretable information which a radiologist can utilize to understand model decision-making and identify cases in which the model may fail to perform. The TAMIL and CACU-Net pipelines have the potential to be added to the screening clinical workflow for a rapid, objective augmentation of radiologist findings. When the COVID-19 pandemic began in 2019 and CT served as a potential method of evaluation for severe COVID-19 patients, the techniques developed here were adjusted for COVID-19 evaluation. Thus, the final aim of this dissertation was to develop a multi-modal model which could aid clinicians in identifying when patients should undergo corticosteroid administration during their course of treatment. This algorithm included 1) a novel segmentation architecture, 2) an investigation of an improved TAMIL algorithm, and 3) comorbidity data. The proposed model demonstrates comparable classification performance compared to the unimodal variants with added interpretability. This technique could improve patient care during future waves of COVID-19, particularly in those patients that are immunocompromised and may require more aggressive treatment. The research provided in these three aims has the potential to improve thoracic CT evaluation by providing more flexible, modality-appropriate models that may augment human readers at various stages of the clinical workflow. The application of such deep learning algorithms has significant potential to enhance clinical efficiency and to ultimately improve patient outcomes.