Knowledge-Guided Framework for Synthesizing Contrast-Dependent Data from Multi-Sequence Non-Contrast MRI

TL;DR

This paper introduces KGSynth, a new method to create contrast-enhanced MRI images from non-contrast scans, improving accuracy and avoiding the need for contrast agents.

Contribution

The novel KGSynth framework integrates knowledge guidance to enhance lesion detail preservation in synthesized MRI images.

Findings

KGSynth achieved an SSIM of 0.567 and PSNR of 19.48 dB for cardiac LGE synthesis.

For brain CBV synthesis, KGSynth yielded an SSIM of 0.697 and PSNR of 24.49 dB.

The method improved accuracy in delineating myocardial infarctions and tumor regions.

Abstract

Background: Contrast-enhanced magnetic resonance imaging (MRI), including late gadolinium enhancement (LGE) and cerebral blood volume (CBV) maps, is essential for characterizing pathologies such as myocardial scars and brain tumors. However, acquiring these images requires gadolinium-based contrast agents (GBCAs), which are contraindicated in certain patient populations. Although deep learning enables cross-modality image translation, current methods frequently fail to preserve lesion details, limiting their clinical utility. Methods: We propose KGSynth, a knowledge-guided framework designed to synthesize contrast-enhanced MRI from non-contrast sequences. This approach incorporates a knowledge estimator to extract lesion and anatomical features, paired with a style mapping network to capture contrast-specific visual characteristics. By explicitly modeling these distinct components, the framework aims to improve pathological fidelity in the synthesized images. Results: Extensive validation on cardiac and brain MRI datasets indicates that KGSynth outperforms existing competing methods. In cardiac LGE synthesis, the model achieved an SSIM of 0.567 and PSNR of 19.48 dB. Similarly, for quantitative brain CBV map synthesis, it yielded an SSIM of 0.697 and PSNR of 24.49 dB. Notably, the method demonstrated improved accuracy in delineating myocardial infarctions and tumor regions compared to baseline models. Conclusions: Integrating explicit knowledge guidance into generative models effectively produces diagnostic-quality images without GBCAs. KGSynth preserves pathological accuracy, offering a viable solution for virtual contrast enhancement. This approach holds promise for clinical workflows, particularly for patients with contraindications to contrast agents.

Figures