TSR-Driven CNN Optimization for Accurate and Interpretable Nonalcoholic Fatty Liver Disease Diagnosis

Abstract

Nonalcoholic Fatty Liver Disease (NAFLD) has emerged as one of the most prevalent chronic liver disorders worldwide, closely associated with sedentary lifestyles, obesity, and metabolic dysfunctions. Early detection is challenging due to the asymptomatic nature of initial stages and variability in imaging quality. Conventional ultrasound-based diagnosis is limited by operator dependency and subjective interpretation, while manual feature extraction and classical machine learning approaches often fail to capture subtle hepatic textural variations, restricting sensitivity in early-stage disease. This study proposes a fully automated, hybrid framework for NAFLD assessment from ultrasound images, integrating Convolutional Neural Networks (CNNs) with Tree-Structured Regularization (TSR) and metaheuristic optimization. CNNs enable hierarchical, data-driven feature extraction, while TSR imposes a biologically inspired hierarchical structure on features, enhancing interpretability and preventing overfitting. Metaheuristic optimization algorithms further fine-tune hyperparameters and select optimal feature subsets, improving both accuracy and model generalization. The framework emphasizes robustness across heterogeneous ultrasound systems, high sensitivity in mild steatosis, and computational efficiency suitable for real-time applications. Experimental evaluations demonstrate that TSR-optimized CNNs outperform traditional optimization methods, achieving higher classification accuracy, faster convergence, and increased resilience to noise. Feature activation analyses indicate improved discriminative representation, confirming the effectiveness of hierarchical optimization in guiding CNN learning. The hybrid framework reduces reliance on invasive diagnostic procedures and supports objective, reproducible, and clinically meaningful assessment of hepatic steatosis.