Bridging synthetic and real images: a transferable domain adaptation method for multi-parametric mapping via multiple overlapping-echo detachment imaging.

Abstract

Objective.This study aims to address the challenge of domain discrepancies between synthetic and real data in quantitative MRI, particularly in multi-parametric mapping using multiple overlapping-echo detachment (MOLED) imaging, which provides rapid and versatile imaging for clinical applications.Approach.A domain adaptation method named MaskedUnet was proposed. Specifically, we employed a mask-based self-supervised pre-training model to learn knowledge from unlabeled real MOLED images. Guided by the learned knowledge of the real data distribution, we regenerated synthetic data closer to the real data distribution to enhance the model's generalization ability to real data. Evaluations were performed onT₂andT₂^*MOLED imaging data of two healthy brain volunteers,T₂and apparent diffusion coefficient (ADC) MOLED imaging data of a healthy brain volunteer, andT₂and ADC MOLED imaging data of 24 patients with brain tumors from 3T MRI scanners were performed, and the results were compared with existing methods to evaluate the effectiveness of the proposed method.Main results.Experimental results demonstrate the effectiveness of the proposed method for MOLED imaging, significantly reducing noise and eliminating streaking artifacts. The normalized mean square error, peak signal-to-noise ratio and structural similarity index of the reconstructed quantitative maps from our method are 0.2170/0.1624, 18.2492/22.7896 dB, 0.7744/0.8162 respectively forT₂/T₂^*of the healthy participants, 0.2423/0.0893, 17.3168/21.9115 dB, 0.7655/0.8416 respectively forT₂/ADC of the healthy participant, and 0.1344, 21.2407 dB, 0.8333 respectively for ADC of the healthy participant.Significance.MaskedUnet demonstrates the feasibility to bridge the gap between synthetic and real MOLED data, advancing the application of multi-parametric MOLED quantitative imaging.