Non-alcoholic fatty liver disease (NAFLD), a prevalent chronic liver condition, demands accurate, non-invasive diagnostics to replace invasive liver biopsies for staging steatosis, non-alcoholic steatohepatitis (NASH), and fibrosis. Deep learning (DL) has demonstrated transformative potential in enhancing diagnostic accuracy and efficiency by leveraging ultrasound (US) imaging, elastography, and clinical/serological data. This systematic review analyzes 64 studies from 2015 to 2025, retrieved from multiple scholarly databases, to evaluate DL models for NAFLD detection and quantification. The reviewed models, primarily leveraging convolutional neural networks (CNNs) and multimodal data integration, achieve high diagnostic accuracy (AUC > 0.90) and generalizability in detecting and staging NAFLD. Ablation studies highlight the critical role of multimodal inputs and advanced architectures in improving performance. However, gaps such as limited diverse datasets, scarce prospective validations, and poor model explainability persist. Opportunities include developing explainable AI (XAI), federated learning for multi-institutional collaboration, and integration with telemedicine for scalable diagnostics. These findings suggest that DL-based systems can significantly reduce biopsy dependency, enhance early detection, and improve clinical outcomes, provided interdisciplinary efforts address existing challenges.