Introduction: Ultrasound (US) is widely used for diagnosing liver disease, particularly cirrhosis, with key signs including liver shape irregularity and echostructure. The classification of liver parenchyma as smooth or coarse, indicative of chronic liver disease, is subjective and dependent on operator experience. To address this, we introduce LivGuard, a deep learning binary classifier designed to detect cirrhosis from a single ultrasound image, applicable to both general and point-of-care (POCUS) USD.
Objective: The objective of the following study was to diagnose cirrhosis by analyzing the liver echostructure through artificial inteligence (AI).
Methods: The dataset was composed of 1625 two-dimensional, ultrasound liver images annotated as cirrhotic (N=677) or not (N=948) captured from 165 individuals at Sanatorio Sagrado Corazon. We stochastically split the master set into training (N=1297; 79.8%), validation (N=159; 9.7%), and test (N=169; 10.2%) sets that were completely disjoint. The output of the efficientNetv2 convolutional neural network (CNN) was an score between 0 and 1 to exhibit the probability of cirrhosis.
Results:The AI system achieved an accuracy in the test set of 88.7%. Sensitivity, specificity, positive (P) and negative (N) predictive values (PV) were 88.8%, 88.5%, 85.5% and 92.2%, respectively. The system was additionally evaluated in a test set of images (N=180; positive for cirrhosis=64) obtained through Butterfly POCUS. The AI system achieved an overall detection rate of 88.8%. Sensitivity, specificity, positive (P) and negative (N) predictive values (PV) were 100%, 82.7%, 76.1% and 100%, respectively.
Conclusion:LivGuard is proven to be a high performer as cirrhosis classifier in ultrasound images. Further work is required to validate this algorithmic framework in prospective cohorts of patients in additional clinical trials and/or real-world datasets.
ALEH Congress 2024: E-POSTER (P-#109)