FINDER: Zero-Shot Field-Integrated Network for Distortion-free EPI Reconstruction in Diffusion MRI

A research team from KAIST and Harvard Medical School has introduced FINDER (Field-Integrated Network for Distortion-free EPI Reconstruction), a novel AI framework that addresses a persistent challenge in diffusion MRI. Echo-planar imaging (EPI) sequences, while fast, are highly sensitive to magnetic field (B₀) inhomogeneities, causing severe geometric distortions that compromise diagnostic accuracy. FINDER reformulates reconstruction as a joint optimization problem, simultaneously estimating both the underlying anatomical image and the distortion-causing field map through an alternating minimization strategy.

At its core, FINDER employs a physics-guided unrolled network architecture that integrates dual-domain denoisers and virtual coil concept extensions to enforce robust data consistency. Crucially, it pairs this with an Implicit Neural Representation (INR) conditioned on spatial coordinates and latent image features. This INR models the off-resonance field as a continuous, differentiable function, allowing the system to effectively disentangle susceptibility-induced distortions from true anatomical structures without requiring pre-training on massive, labeled datasets.

The zero-shot, scan-specific approach means FINDER can be applied immediately to individual patient scans, adapting to unique acquisition conditions and field imperfections. Experimental results demonstrate superior geometric fidelity and image quality compared to state-of-the-art baselines. This represents a significant advancement over traditional methods that either require separate, time-consuming field mapping scans or struggle to integrate distortion correction seamlessly into the reconstruction pipeline.