ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


ERS2319. Artificial Intelligence for Detection of Nonalcoholic Fatty Liver Disease on B-Mode Ultrasound
Authors * Denotes Presenting Author
  1. Aylin Tahmasebi *; Thomas Jefferson University
  2. Frank Yang; Thomas Jefferson University
  3. Corrine Wessner; Thomas Jefferson University
  4. Shuo Wang; Thomas Jefferson University
  5. Flemming Forsberg; Thomas Jefferson University
  6. Flavius Guglielmo; Thomas Jefferson University
  7. John Eisenbrey; Thomas Jefferson University
Objective:
Non Alcoholic Fatty Liver Disease (NAFLD) is the most common cause of liver disease in United States. Liver biopsy is the gold standard in diagnosing NAFLD and the most accurate tool for grading fibrosis, however is invasive and carries the risk of complications. Ultrasound, as a non-invasive and readily available tool, has an important role in diagnosing NAFLD, however it underestimates the prevalence of hepatic steatosis in mild cases. This prospective study aimed to evaluate the use of a deep learning-based model to diagnose NAFLD on B-mode ultrasound.

Materials and Methods:
As part of an ongoing IRB-approved study, patients scheduled for liver fat quantification via proton density fat fraction on magnetic resonance imaging (MRI) were recruited and consented. The exclusion criteria were: significant alternate diagnosis for chronic liver disease, history of liver transplant, or right upper quadrant malignancy or focal liver lesion larger than 5 cm. Ultrasound imaging was performed using a Logiq E10 scanner with a C1-6 probe (GE Healthcare, Waukesha, WI). Images from 10 different locations in the right and left hepatic lobe were collected following a standard protocol. All images were de-identified and cropped using an automated cropping script written in Matlab (2016a, The MathWorks Inc., Natick, MA). A deep learning algorithm was trained using all the images while subject recruitment is ongoing for a prospective testing cohort. Diagnostic accuracy was established using ROC analysis. MRI-based liver fat quantification was used as the reference standard for NAFLD with =6.4% considered normal and >6.4% indicative of NAFLD.

Results:
A deep learning-based anomaly detection model and optimal scoring algorithm were developed using the combination of two Tensor Flow Keras models of generative adversarial networks (GAN) generator and discriminator for NAFLD detection. 1063 images from 102 subjects were used for model training. Within this training set, the model achieved an AUC of 0.89 based on the ROC analysis with 92% average precision on 2-classes Precision-Recall chart. Importantly, this performance exceeds prior radiologist performance in a similar dataset (81% sensitivity and 86% specificity).

Conclusion:
Preliminary anomaly detection results using a GAN deep learning approach are encouraging, indicating it may be useful as a screening tool for NAFLD on B-mode ultrasound.