Speckle-Aware Hybrid Learning for Fatty Liver Assessment from B-mode Ultrasound

Description

Abstract—B-mode ultrasound is widely used for screening and
monitoring fatty liver (metabolic dysfunction–associated steatotic
liver disease, MASLD), but speckle noise, operator dependence and
scanner variability limit reliable grading. This paper (i) summarises
common ultrasound enhancement, segmentation and machine learning
approaches for liver assessment and (ii) proposes a practical, speckle
aware hybrid framework (SA-LiverNet) combining liver ROI
localisation, attention-based segmentation, and fusion of radiomics
with deep embeddings for robust grading. We also outline a
reproducible evaluation protocol (dataset reporting, metrics, ablations,
calibration and external validation) to bridge prototype pipelines and
clinically credible systems. A qualitative demonstration using
representative ultrasound images illustrates preprocessing and region
level feature analytics.
Index Terms—Ultrasound imaging, MASLD/NAFLD, fatty liver,
speckle noise, liver segmentation, attention U-Net, radiomics,
explainable AI.