UC Riverside

ASL is a magnetic resonance imaging technique used to assess cerebral blood flow by magnetically labeling incoming blood. Instead of relying on external agents, ASL takes advantage of the water in arterial blood as an endogenous tracer. One of the major concerns with ASL imaging is the low Signal-to-Noise Ratio (SNR). Deep learning models can be employed to enhance the quality of ASL images, thus paving the way for more accurate clinical diagnoses. We use data from 63 subjects, divide them into training, validation, and testing data to train out four deep learning models - UNET 2D, UNET 3D, VisionTransformer2D, VisionTransformer3D. All the models improved the input images significantly but UNET 3D showed the highest improvement. Considering less data requirement, and time efficiency, UNET 3D outperforms all models. All the results were verified with 5-fold cross-validation.