Pretext Task Adversarial Learning for Unpaired Low-field to Ultra High-field MRI Synthesis

TL;DR

This paper addresses the scarcity of high-field MRI data by proposing an unpaired low-field to ultra high-field MRI synthesis method. The method, called Pretext Task Adversarial (PTA) learning, integrates Slice-wise Gap Perception (SGP) using contrastive learning, Local Structure Correction (LSC) via a pretext task of rotating and masking local blocks, and adversarial training guided by a pretext task to improve realism. The framework computes a composite objective $L_{all} = \lambda_1 \mathcal{L}_{syn} + \lambda_2 \mathcal{L}_{cycle} + \lambda_3 \mathcal{L}_{adv}$ with weights $\lambda_1=0.5$, $\lambda_2=0.2$, $\lambda_3=0.3$, and uses a discriminator to enforce realism, while a cycle-consistency constraint preserves anatomical fidelity. Extensive experiments demonstrate state-of-the-art performance on low-field to ultra high-field MRI synthesis, achieving $16.892$ for FID, $1.933$ for IS, and $0.324$ for MS-SSIM, and ablations plus MRI physicist evaluation support the contributions.

Abstract

Given the scarcity and cost of high-field MRI, the synthesis of high-field MRI from low-field MRI holds significant potential when there is limited data for training downstream tasks (e.g. segmentation). Low-field MRI often suffers from a reduced signal-to-noise ratio (SNR) and spatial resolution compared to high-field MRI. However, synthesizing high-field MRI data presents challenges. These involve aligning image features across domains while preserving anatomical accuracy and enhancing fine details. To address these challenges, we propose a Pretext Task Adversarial (PTA) learning framework for high-field MRI synthesis from low-field MRI data. The framework comprises three processes: (1) The slice-wise gap perception (SGP) network aligns the slice inconsistencies of low-field and high-field datasets based on contrastive learning. (2) The local structure correction (LSC) network extracts local structures by restoring the locally rotated and masked images. (3) The pretext task-guided adversarial training process introduces additional supervision and incorporates a discriminator to improve image realism. Extensive experiments on low-field to ultra high-field task demonstrate the effectiveness of our method, achieving state-of-the-art performance (16.892 in FID, 1.933 in IS, and 0.324 in MS-SSIM). This enables the generation of high-quality high-field-like MRI data from low-field MRI data to augment training datasets for downstream tasks. The code is available at: https://github.com/Zhenxuan-Zhang/PTA4Unpaired_HF_MRI_SYN.

Figures (5)

• Figure 1: The motivation and challenges of our PTA framework. (a) Clinical motivation lies in synthesizing high-field MRI from low-field MRI. (b) The challenge is the clinical fidelity gap between low-field MRI data and high-field MRI data.
• Figure 2: Workflow of the proposed pretext task adversarial (PTA) learning framework. (a) Pretext Task Pre-training focuses on reducing domain gaps between low-field and high-field MRI through slice-wise gap perception and local-structure correctness. (b) Data Synthesis and Adversarial Training aims to synthesize high-field MRI data from low-field MRI inputs using a synthesis network.
• Figure 3: Visualization result of SGP and LSC process. (a) SGP enhances inter-slice similarity by pre-training on sequential low-field and high-field MRI data and matching the most similar slices within a randomly shuffled batch. (b) LSC enhances local structures by recovering fine image details from locally masked and rotated images.
• Figure 4: Visualization result of different baselines for exemplar regions (top and bottom row of each panel) (a) Synthesis of high-field MRI data from cross-contrast low-field MRI data. (b) Synthesis of ultra high-field MRI from same-contrast low-field MRI.
• Figure 5: Ablation study and MRI physicist evaluation of our PTA.(a) Performance comparison of different PTA configurations through an ablation study. (b) MRI physicist evaluation study, comparing quality scores between low-field MRI and synthesized MRI using a 5-point Likert scale. (c) MRI physicist evaluation statistics, including box plots, histograms, and correlation analysis.