Executive Summary : | Magnetic Resonance (MRI) Imaging is a non-invasive method used in modern radiology to assess anatomical regions and pathological processes. Machine learning, particularly deep learning, plays a crucial role in reducing workflow complexity, overcoming inexperience due to improper acquisition protocols, and increasing throughput with high-quality workflows and time-efficient imaging methods. However, current deep learning models for MR imaging methods require discrete instantiation and re-training for every configuration of acquisition settings under a low training data regime. This can lead to a proliferation of models for each combination of settings, making workflow maintainability and configuration management challenging. Task-specific models may not interpolate or extrapolate in the imaging task context, making quick adaptation to new tasks difficult. Additionally, large amounts of training data are needed to model reusable representations, making preparation a demanding exercise. Meta-learning is an emerging machine learning paradigm that aims to address these limitations by absorbing knowledge from various MRI acquisition contexts and using the same to generalize to unseen tasks proficiently. The focus of this research is building a meta-learning model for regression tasks like image reconstruction and image-to-image translation problems using real-world multi-modal MRI data. The approach is incorporated on well-known problems like MRI reconstruction, super-resolution, and image translation, as well as a novel application like Dynamic Contrast Enhanced MRI for datasets of complex anatomy like prostate. Experiments with clinical relevance on MRI reconstruction show that this approach performs much better than conventional deep learning models, opening pathways to devise a smarter MRI workflow. |