RGB2qMRI: Can Deep Learning Models for Quantitative MRI be Trained with RGB Pictures?

Motivation: The shortage of diverse ground-truth quantitative MRI (qMRI) maps limits the development of deep learning-based techniques for qMRI. Goal(s): Our aim is to use general color pictures to train deep learning models that are capable of qMRI reconstruction. Approach: To create a training dataset, the red, green and blue channels of color pictures are used as quantitative maps, which are employed to synthesize weighted images through a MR physics forward model. Using QTI as example, the trained deep learning models are validated with in vivo data. Results: The evaluation demonstrates that the models trained on RGB pictures can generalize to in vivo data. Impact: We demonstrated the feasibility of using widely available, non-anatomical RGB pictures to train deep learning models for qMRI, addressing the issue of limited training data, reducing costs of building training datasets, and avoiding biases toward specific health or disease types.