Self-supervised deep learning model for denioising and distortion correction in accelerated echo planar imaging

Motivation: Accelerated EPI can improve temporal resolution but leads to low SNR, which exacerbates geometry distortions and aliasing artifacts. Goal(s): Our objective is to simultaneously denoise and correct distortions in accelerated EPI using self-supervised learning. Approach: An raw image generation of accelerated EPI block reproduces distorted images from distortion-corrected output of the network. The model is trained to reduce the error between the original distorted image and the synthesized image, enabling effective artifact correction with denoising based on J-invariant characteristic. Results: Even in the absence of ground truth images, our model effectively corrects aliasing artifacts and geometric distortions in accelerated EPI images, providing denoising. Impact: The proposed self-supervised approach achieves a significant advancement by enabling simultaneous denoising and distortion correction in accelerated EPI without ground truth images, thereby enhancing image quality in accelerated imaging.