Using a deep residual learning framework to enhance APT-weighted (APTw) MR images of brain tumors acquired by SENSE with compressed sensing

Sensitivity encoding (SENSE) is often adopted to accelerate image acquisition for various MRI sequences, including APTw. This is further accelerated with SENSE with compressed sensing (called CS-SENSE), but the image quality degrades to some extent. We collected both SENSE- and CS-SENSE-APTw images and trained a generative model with residual learning to generate SENSE images from CS-SENSE images. The generated results were proved to be highly similar to SENSE-APTw images and less noisy than both SENSE- and CS-SENSE-APTw images. With a larger dataset, we can train more robust models and eventually replace SENSE- with CS-SENSE for a speedup of ~50%.