Predictive Modeling for Non Alcoholic Fatty Liver Disease Detection

Non-Alcoholic Fatty Liver Disease (NAFLD) is a prevalent liver disorder globally, with potential progression to severe complications such as cirrhosis if left untreated. This study leverages deep learning and explainable AI (XAI) tech- niques to classify and interpret ultrasound images for NAFLD detection. We train both CNN (EfficientNetB0) and Transformer (DeiT) models, achieving a peak accuracy of 95.12% with DeiT. Additionally, a feature-engineering approach is employed, extracting image texture features—such as contrast, energy, and homogeneity—and integrating them with image feature vectors to enhance machine learning classifiers. SHAP analysis revealed that these features played a crucial role in classification, underscoring their clinical relevance. Traditional machine learning models trained on combined features demonstrated strong performance, with XGBoost emerging as the best performer (95% accuracy). To further improve classification, we explore multi-task learning by pre-training on fibrosis staging data, achieving a fibrosis classification accuracy of 97.63% and a slight drop in NAFLD classification accuracy (92.68%) post fine-tuning. Dimensionality reduction techniques like PCA enhance computational efficiency. Emphasizing explainability, SHAP provides insights into model decision-making, bridging the gap between clinical interpretation and AI-driven predictions. This research highlights the potential of deep learning and hybrid feature-engineering approaches in advancing NAFLD diagnosis and staging.