Skip to main content
ACRADIO 2023 Annual Meeting Posters Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Informatics and Data Science

Lesion Detection Using Artificial Hepatobiliary Phase Post-Contrast Images Created with a Generative Adversarial Network

Saturday, May 6, 2023

    First Author(s)

  • VP

    Vincent G. Parenti, MD

    Resident Physician
    University of Texas Southwestern Medical Center
    Dallas, Texas, United States

    • Co-Author(s)

  • CB

    Chandan Ganesh Bangalore Yogananda, PhD

    Assistant Professor and Imaging Deep Learning Scientist
    University of Texas Southwestern Medical Center, United States

  • YN

    Yee S. Ng, MD

    Assistant Professor of Radiology
    University of Texas Southwestern Medical Center, United States

  • AT

    Ali Tejani, MD

    Resident Physician
    University of Texas Southwestern Medical Center, United States

  • IP

    Ivan Pedrosa, MD, PhD

    Professor of Radiology and Vice Chair of Radiology Research
    University of Texas Southwestern Medical Center, United States

  • JR

    Jesse Rayan, MD

    Assistant Professor of Radiology
    University of Texas Southwestern Medical Center, United States

Purpose: Hepatobiliary specific gadolinium based contrast agents are used to improve liver lesion detection and characterization at the expense of increased MRI scanning duration. We explored using generative adversarial networks (GAN) to create artificial hepatobiliary phase (HBP) images from pre-contrast T1-weighted images (pre-T1). This has the potential to increase sensitivity for lesion detection on non-contrast and non-hepatobiliary specific protocols, as well as reduce MRI scanning time.

Methods/Materials: 10 studies utilizing a hepatobiliary MRI contrast agent were selected at random. A Super-Resolution GAN was trained using an axially aligned pair of pre-T1 images and HBP images from each study. 7 studies were used for training with standard augmentation techniques on 128x128 pixel cropped patches, and 2 studies were used for validation using full sized 512x512 pixel image inputs. A trained radiologist subjectively compared the artificial and true HBP images for resolution, artifact, and lesion assessment.

Results: GAN generated artificial HBP images were of similar resolution compared to the true HBP images, and they were of higher resolution compared to the pre-T1 images. The true HBP images are of higher resolution (i.e. thinner slices) than the pre-T1 dynamic 3D GRE sequence images. Artificial HBP image artifact was within acceptable limits. None of the artificial HBP images contained hyperintensity within the biliary system which is seen in the true images. Finally, lesion conspicuity and enhancement features were similar between the artificial and true HBP images.

Conclusions: GANs have the potential to produce high resolution artificial HBP images from pre-T1 images, which can aid in lesion detection.