ScholarWorks@성균관대학교

초록

T1-weighted (T1w) and T2-weighted (T2w) magnetic resonance imaging (MRI) are the most widely used structural imaging modalities, providing complementary information on brain morphology. However, acquiring both types of imaging data is resource-intensive and time-consuming, particularly for populations prone to significant head motion, such as young children and the elderly. One possible solution is image synthesis. Recent advancements in deep learning-based models enabled the synthesis of T1w and T2w MRI. However, few studies have explored synthesis models incorporating age information spanning the entire lifespan, from early development to advanced age. In this study, we proposed a deep learning-based cross-modal synthesis model for T1w and T2w MRI. Our model achieved high similarity between actual and synthesized images, with mean squared error values of 0.024, 0.022, and 0.024, and peak signal-to-noise ratio values of 20.26, 21.13, and 18.96 for early development, young adulthood, and elderly cohorts, respectively. Furthermore, we demonstrated reliable age-related effects in the synthesized data, showing decreases in microstructure profile covariance gradient values in the medial and lateral prefrontal cortex and visual cortex, as well as cortical thinning in the sensory/motor regions, precuneus, frontoparietal regions, and limbic areas. Clinical Relevance-The developed pipeline can generate multimodal MRI data for participants facing challenges during MRI scanning.

키워드