Turing Testing the Realism of U-Net, GAN, and Real Images for Virtual Contrast-Enhanced Breast MRI

Motivation: Contrast agent-based breast MRI is limited by cost, time, and contraindications for certain patients. Goal(s): Assess the realism of virtual contrast-enhanced (vCE) images compared to real contrast-enhanced images. Approach: Compare GAN and U-Net models in generating vCE images from unenhanced sequences, evaluating results with quantitative metrics and a Turing test. Results: The U-Net outperformed the GAN on quantitative metrics. The Turing test showed similar classification rates for real and GAN images within each reader, though there were disagreements between readers. In contrast, U-Net images showed greater variability, with some readers rating them as real at notably high or low rates. Impact: DL-based vCE provides an effective alternative to significantly reduce reliance on contrast agents, addressing concerns of cost, time, and contraindications. Incorporating adversarial training enables the model to learn intricate details of contrast-enhanced images, resulting in more realistic outputs.