Reproduce “HepatoGEN: Generating Hepatobiliary Phase MRI with Perceptual and Adversarial Models” on OpenResearch

About this paper

Jens Hooge, Gerard Sanroma-Guell, Faidra Stavropoulou, Alexander Ullmann, Gesine Knobloch, Mark Klemens, Carola Schmidt, Sabine Weckbach and 1 more

Researchers at Bayer AG developed and compared deep learning models to synthesize hepatobiliary phase (HBP) MRI images from earlier contrast phases, aiming to reduce liver MRI scan times. The study found that a perceptual GAN (pGAN) achieved the best quantitative metrics, while a perceptual U-Net (UNet_mid) provided a favorable balance in qualitative assessments, demonstrating potential for reduced scan times and improved diagnostic utility in some cases.

Key insights

• The Contrast Evolution Score (CES) effectively curates training data, improving the quality of synthetic images and enabling models to generate clearer liver features, even for challenging cases with suboptimal real HBP images.
• There is a discrepancy between quantitative image quality metrics and expert subjective clinical preferences; pGAN scored best quantitatively, but UNet_mid often balanced artifact suppression and anatomical detail more favorably in qualitative reviews.
• Deep learning-based HBP image synthesis can potentially enhance diagnostic utility for cases with suboptimal real HBP images by improving liver contrast and structural detail in the synthetic output.

Summary by alphaXiv.