Breast cancer is found in one in eight women in the United States and is expected to be the most frequently diagnosed form of cancer among them in 2018. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) plays a significant role in breast cancer management, especially for high-risk patients. The interpretation of the DCE-MR images remains labor-intensive and can lead to erroneous clinical results with subsequent unnecessary biopsies and patient stress. Computer-aided detection (CADe) and diagnosis (CADx) systems, or radiomics, have been developing to help reduce these errors. The conventional CADx methods involve automatic segmentation of a lesion from the neighboring background and extraction of intuitive hand-crafted features, previously developed by the engineers and domain experts. Such features describe lesion’s size, shape, texture, and enhancement patters. The recent advances in machine learning techniques have provided an alternative method for image assessment, where images are analyzed directly by deep learning models. ,Radiomcs has strong potential to lead clinicians towards more accurate and rapid image interpretation. Furthermore, it can serve as a “virtual digital biopsy”, allowing for the discovery of relationships between radiomics and the pathology/genomics from actual biopsies, for ultimate use when biopsies are not possible. The objective of this research is to analyze the conventional and design new radiomics methods for breast DCE-MRI, in order to improve image-based clinical decisions. Specifically, part of the dissertation studies the robustness of conventional radiomics methods across MRI scanners of different manufacturers. Another focus of the research is developing accurate and robust deep learning-based models for automated breast lesion characterization, tailored to the complex 4-dimentional (4D) DCE-MRI data. These models are applied to two clinical tasks, lesion malignancy assessment and prediction of cancer response to therapy. The research is concluded by testing the ultimate hypothesis that incorporating the two types of radiomics, deep learning-CADx and conventional-CADx, will enhance lesion characterization within the tasks of diagnosis and treatment response.,The research presented the following results. First, the robustness analysis revealed radiomics features that are generalizable across datasets acquired with MRI scanners of two major manufacturers. Specifically, features that characterize lesions in terms of size are robust in their average values. Furthermore, entropy feature, which quantifies randomness of pixel values of the lesion image, is robust in its classification performance in multiple clinical tasks. Second, a novel deep learning-based method was developed to assess breast lesion malignancy and response to therapy based on the DCE-MRI sequence. Finally, the results demonstrated that deep learning-based methods are complimentary to the conventional radiomics. ,The medical significance of this research is that it has potential to improve DCE-MRI-based breast cancer management. The developed deep learning methods and their fusion with conventional radiomics can reduce human burden and allow for more rapid and accurate analysis of the images.