In the past few years, applications of deep learning have experienced explosive growth due to their role in solving complex problems. Deep learning has recently been gaining attention for use in medical imaging and applications of deep learning are being explored to enhance radiology practice, including for the selection and preparation of images for interpretation, analysis of image quality, and assistance for diagnostic decision-making tasks, among many other clinical applications. For the use of deep learning in medical imaging, it is important to understand physical limitations of medical images as well as techniques with which to augment inputs and forms of output with which to enhance specific task performance. The primary goals of this research are to investigate deep learning in medical imaging through contributions in (i) workflow enhancement, (ii) diagnostic improvements, and (iii) AI output understanding (i.e., explainability) through the specific tasks of detection and visualization of pneumothorax on thoracic radiographs. However, this specific investigation of pneumothorax detection and visualization could yield procedures applicable to other imaging applications. Pneumothorax, the abnormal presence of air between the lung and chest wall, can be diagnosed using a chest radiograph; visual indications of pneumothorax in a chest radiograph include a fine line at the edge of the lung and a change in texture outside the lung. Due to the overlapping structures within a frontal chest radiograph due to 2D projection radiography, pneumothorax can be difficult for even an experienced radiologist to detect. The detection of pneumothorax within the radiograph is further complicated by the wide variety of sizes and severities pneumothorax can possess. This work demonstrates the potential applications for deep learning to medical imaging tasks including the enhancement of radiology workflow, improving medical image diagnosis, and explanatory output from deep learning algorithms. Workflow enhancement can be achieved through the use of a deep learning model for the classification of radiographic views from a dual-energy, a standard, or a portable chest radiography study, reducing the reliance on DICOM headers for proper display and storage. Deep learning can improve diagnosis through the detection of pneumothorax on frontal chest radiographs; this work demonstrated the impact of the effective input image resolution on deep learning performance, indicating the importance of deep learning algorithms customized for the task being performed. Human-interpretable and explanatory output from deep learning algorithms is needed for clinical implementation. This work showed visualizations of pneumothorax detection on the images and quantified the performance of the visualizations. Overall, this work demonstrates the potential of deep learning to be applied in radiology practice to enhance workflow, improve and enhance diagnosis of medical images, as well as provide human-interpretable explanations of the output.