Deep Learning Meets Graph: Exploring Mutually Reinforcing Paradigms for Improved Quantitative Medical Image Analysis