Motivation: Susceptibility artifacts in EPI are a barrier to image quality and anatomical accuracy, particularly challenging in multi-echo data due to varying distortions across separate echo times. Goal(s): Our goal is to devise a physics-driven, deep unsupervised model for efficient and robust artifact correction in multi-echo EPI scans. Approach: We developed a multi-echo Forward-Distortion network (meFD-Net) with parallel decoding pathways to process multi-echo sequences, synthesizing reversed-phase-encode pairs per echo for unsupervised learning and field estimation. Results: meFD-Net achieves competitive artifact correction with the gold-standard TOPUP method, while significantly shortening processing time and demonstrating high fidelity to anatomical structures in multi-echo EPI. Impact: Applicability on multi-echo EPI scans, high anatomical accuracy and fast processing times enabled by meFD-Net significantly enhances feasibility in clinical and research settings. These advancements can facilitate real-time EPI applications through efficient, artifact-free imaging across diverse conditions.
Alkilani et al. (Tue,) studied this question.