Pixel-level transformer GAN for enhanced parametric mapping of DCE MRI analysis

Background

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) plays a crucial role in the diagnosis and monitoring of cancers, as it reveals physiological and vascular characteristics of tumors. Traditional pharmacokinetic modeling necessitates high temporal resolution, resulting in relatively low signal-to-noise ratio (SNR) and spatial resolution with limited allocated time for each phase.

Purpose

To explore the feasibility of using deep learning with sparse DCE MRI phases to generate dense temporal resolution DCE-MRI-derived parametric map.

Methods

An innovative approach, the vision transformer Pix2Pix generative adversarial network (VP-GAN), was introduced to translate the sparse DCE-MRI series into dense-phase DCE-MRI-based parametric maps, specifically targeting K^trans and v_e. The strengths of both Vision Transformers and GANs were utilized to capture complex temporal dynamics and spatial features. The proposed method was comprehensively compared with several existing deep learning models, both for the entire image and within regions of interest (ROI). Metrics used for comparison included Peak-Signal-to-Noise-Ratio (PSNR), Structural Similarity Index (SSIM), Pearson correlation analysis, and Bland-Altman analysis. Additionally, ROI histogram analysis was performed to assess the distribution of parametric values.

Results

The parametric maps generated by the proposed approach were qualitatively and quantitatively consistent with the reference images. The performance of the comparative studies evidenced the superiority of VP-GAN over other approaches.

Conclusion

The proposed model performs well in converting DCE-MRI with a subset of uniformly spaced time points into physiological parametric maps derived from dense-phase DCE-MRI, allowing for DCE-MRI analysis with much fewer phases.