Self-supervised multi-instance contrastive learning for reduction of cardiac bSSFP off-resonance artifacts

Motivation: Off-resonance artifacts in bSSFP cine images preclude the assessment of cardiac function. Acquiring corrupted-clean image pairs for supervised methods is challenging. Goal(s): To develop a self-supervised adversarial framework to reduce off-resonance artifacts without corrupted-clean image pairs. Approach: We employ multi-instance contrastive learning during the generator's training to enforce point-wise consistency between outputs of artifact-affected bSSFP images with altered RF phase increments. Additionally, we design a rectangle-window spatial-temporal transformer for longer-range dependencies without increasing computational complexity. Results: Experiments on ACDC and in-house datasets show that our method outperforms existing unpaired self-supervised approaches and generates comparable results against several fully-supervised models. Impact: Obtaining corrupted-clean bSSFP image pairs is challenging, particularly with cardiac devices or high-field MR. Our proposed self-supervised approach mitigates off-resonance artifacts and provides a practical solution for reliable bSSFP cine imaging.