Abdominal Ultrasound Similarity Analysis for Quantitative Longitudinal Liver Fibrosis Staging

Non-alcoholic fatty liver disease is one of the most common diffuse liver diseases worldwide, affecting approximately 25–30% of the global population. Accurate liver fibrosis staging and longitudinal tracking are crucial for effective care. Core needle biopsy is the current gold standard but it is invasive, inaccurate, and costly. Abdominal B-mode ultrasound (US) in conjunction with shear wave elastography (SWE) offers a non-invasive alternative but suffers from high variability, impairing quantitative assessments. Consistent liver views in the abdominal ultrasound during multiple visits will likely result in consistent SWE-based fibrosis staging. This study presented a proof-of-concept pipeline for identifying the most anatomically similar transverse view in the right liver lobe in abdominal US videos across multiple visits. Our three-stage framework consisted of liver view classification, liver capsule and hepatic vessel segmentation, and quantitative similarity measurement. We used a pretrained EfficientNet-B3 network for liver view classification, achieving 99% accuracy. We then used Efficient-UNet to segment the liver capsule and hepatic vessels, obtaining Dice scores of 90% and 58%, respectively. The classification and segmentation outputs were used for similarity analysis. We evaluated four similarity metrics including deep image structure and texture similarity (DISTS), Root Mean Squared Error (RMSE), Normalized Cross Correlation (NCC), and structural similarity index measure (SSIM), with SSIM resulting in the best results. This pipeline has the potential to improve SWE-based quantitative fibrosis staging and enable cost effective longitudinal tracking.