Speckle-Robust Local Phase and Ternary Texture Encoding (SLaP-TEX) based Feature Extraction for Liver Steatosis Classification in Ultrasound Imaging

Abstract

Ultrasound imaging is a preferred modality for non-invasive liver steatosis screening, yet the inherent speckle noise and texture ambiguity hinder automated diagnostic precision. Existing convolutional neural networks (CNNs) primarily rely on intensity-based texture cues, overlooking phase-based structural continuity that remains stable under speckle corruption. This study proposes a Speckle-Robust Local Phase and Ternary Texture Encoding (SLaP-TEX) model that combines local phase symmetry descriptors with ternary pattern encoding to generate robust representations from liver ultrasound images. The proposed model enhances boundary localization and fine-grained tissue discrimination through a two-stage encoding pipeline comprising Local Phase Filtering (LPF) and Adaptive Ternary Encoding (ATE). The fused phase-texture maps are processed through a MobileNetV3-Small backbone, offering computational efficiency for real-time deployment. Experiments on the RGM-augmented ultrasound dataset demonstrate superior performance with 99.02 % accuracy, 0.998 AUC, and 0.018 loss, outperforming existing models while maintaining a 2.1 M parameter footprint. The SLaP-TEX model offers a compact, phase-aware, and speckle-resilient feature extractor for clinical ultrasound analytics.