Cross-Contrast Enhancement of Brain Tumor Datasets using Deep Learning

Motivation: Under-sampling is an effective way to reduce MRI acquisition times, but highly under-sampled MRIs lack details required for diagnosis. Goal(s): We enhance under-sampled MRIs using Deep Learning models that leverage complementary information from another contrast, minimizing differences between enhanced images and their fully-sampled counterparts. Approach: We train Dense UNets on a dataset containing synthetically under-sampled MRIs of 369 patients with brain tumors and perform quantitative analysis on both the whole brain and tumor tissues. Results: Models trained with cross-contrast priors outperform those using only under-sampled images and maintain higher fidelity as acceleration increases. Tumor reconstruction errors are higher than in the whole brain. Impact: Leveraging complementary information from another contrast helps overcome the image fidelity lost at higher acceleration factors, allowing for faster diagnostically useful scanning. We extend the analysis of multi-contrast MRI enhancement to patients with tumors, increasing clinical relevance of results.