Magnetic Resonance Imaging (MRI) is one of the preferred non-invasive diagnostic modalities capable of producing complementary contrast images. However, the scan duration is typically long and is accompanied by various undesired interference from patient and acquisition settings. The resulting measurements from the MRI scanner are degraded in quality and the diagnostic utility of the scan is deteriorated. Recently, deep learning has shown promising results in image restoration. However, methods of deep learning have certain limitations. To begin with, the existing models that aim to suppress only a single type of artifact cannot be scaled in the healthcare system. Moreover, current methods do not utilize various nuances of artifacts to formulate objective functions that aid restoration in unseen artifacts and composite artifacts. Second, contemporary methods for disease detection identify healthy tissues as anomalous regions. Third, existing unsupervised methods are not investigated into their properties that influence the models’ capability to detect disease and anomalous artifacts.

In this work, methods to address the limitations mentioned above are developed. First, a curriculum based meta learning (CMAML) for MRI image restoration is proposed that efficiently adapts to several artifacts within a single model. Moreover, the nested bi-level optimization in the proposed method promotes latent representations that characterize knowledge shared across different artifact types. Second, a kernel correction, a post-processing method that modulates the high frequency content of the MRI images for a reliable residue-based disease localization in brain tissue is proposed. Third, four representational properties to reveal the anomaly detection capabilities of four unsupervised methods, along with their relation to the models’ training formulation is presented.