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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cancer Diagnostics

Background:

• Dynamic Contrast-Enhanced MRI (DCE-MRI) is vital for cancer diagnosis, revealing tumor vascularity.
• Traditional DCE-MRI requires high temporal resolution, leading to lower signal-to-noise and spatial resolution.
• Limited time allocation per phase restricts detailed analysis in conventional DCE-MRI.

Purpose of the Study:

• To assess the feasibility of deep learning for generating dense temporal resolution DCE-MRI parametric maps from sparse data.
• To explore the potential of AI in improving the efficiency and quality of DCE-MRI analysis.
• To overcome the limitations of traditional DCE-MRI by leveraging sparse data acquisition.

Main Methods:

• A Vision Transformer Pix2Pix Generative Adversarial Network (VP-GAN) was developed to translate sparse DCE-MRI series into dense-phase parametric maps (K^trans, v_e).
• The model integrates Vision Transformers and GANs to capture intricate temporal dynamics and spatial features.
• Performance was evaluated against existing deep learning models using PSNR, SSIM, Pearson, and Bland-Altman analyses, including ROI histogram analysis.

Main Results:

• VP-GAN generated parametric maps that were qualitatively and quantitatively consistent with reference images.
• The proposed VP-GAN model demonstrated superior performance compared to other evaluated deep learning approaches.
• Comparative analyses confirmed the effectiveness of the model in reconstructing detailed imaging features.

Conclusions:

• The developed model successfully converts sparse DCE-MRI data into physiological parametric maps from dense-phase DCE-MRI.
• This approach enables robust DCE-MRI analysis using significantly fewer imaging phases.
• The findings support the use of AI for more efficient and effective cancer imaging and monitoring.