Deep learning for multi-contrast MRI synthesis

Abstract

Magnetic resonance imaging (MRI) possesses the unique versatility to acquire images under a diverse array of distinct tissue contrasts. Multi-contrast images, in turn, better delineate tissues, accumulate diagnostic information, and enhance radiological analyses. Yet, prolonged, costly exams native to multi-contrast pro-tocols often impair the diversity, resulting in missing images from some contrasts. A promising remedy against this limitation arises as image synthesis that recovers missing target contrast images from available source contrast images. Learning-based models demonstrated remarkable success in this source-to-target mapping due to their prowess in solving even the most demanding inverse problems. Main-stream approaches proposed for synthetic MRI were typically subjected to a model training to perform either one-to-one or many-to-one mapping. One-to-one models manifest elevated sensitivity to detailed features of the given source, but they perform suboptimally when source-target images are poorly linked. Meanwhile, many-to-one counterparts pool information from multiple sources, yet this comes at the expense of losing detailed features uniquely present in cer-tain sources. Furthermore, regardless of the mapping, they both innately demand large training sets of high-quality source and target images Fourier-reconstructed from Nyquist-sampled acquisitions. However, time and cost considerations put significant challenges in compiling such datasets. To address these limitations, here we first propose a novel multi-stream model that task-adaptively fuses unique and shared image features from a hybrid of multiple one-to-one streams and a single many-to-one stream. We then introduce a novel semi-supervised learning framework based on selective tensor loss functions to learn high-quality image synthesis directly from a training dataset of undersampled acquisitions, bypass-ing the undesirable data requirements of deep learning. Demonstrations on brain MRI images from healthy subjects and glioma patients indicate the superiority of the proposed approaches against state-of-the-art baselines.