From Segmentation to Disease Prediction: A Systematic Review of AI Methods for Adipose Tissue Analysis in Medical Imaging

Accurate segmentation and classification of adipose tissue are critical tasks in the field of medical imaging, as they contribute to enhanced medical diagnostics. However, manual segmentation and classification are time-consuming and prone to variability. These limitations drive the development of automated and efficient artificial intelligence (AI) methods, including machine learning (ML) and deep learning (DL). A wide range of ML and DL techniques now exist, offering diverse strategies for medical image analysis and early disease detection. Given this growing body of work, a systematic literature review is necessary to synthesize evidence, compare methodologies, and identify research gaps. Existing literature reviews on adipose segmentation and classification often focus on specific imaging modalities or limited disease applications. Nevertheless, none provide a systematic quantitative evaluation of ML and DL methods. Such a systematic quantitative evaluation is essential for a fair comparison of methodologies and guiding the development of more robust AI models. Moreover, existing reviews overlook key challenges, including limited annotated datasets, feature extraction, and predictive modeling for disease risk. This systematic literature review investigates ML and DL techniques for adipose tissue segmentation, classification, and disease risk prediction. An extensive search across four major scientific databases is conducted, followed by rigorous selection criteria that yield forty high-quality studies. The selected studies are classified into two types: ML-based segmentation and classification, covering fourteen studies, and DL-based segmentation and classification, covering twenty-six studies. Within the ML-based approaches, further analysis focuses on four major radiomic features, providing deeper insights into feature extraction techniques. In contrast, among the twenty-six DL-based studies, nine specifically address limited annotated data, with a focus on semi-supervised learning strategies and data augmentation methods. The review also presents a critical analysis of dataset constraints as well as the inherent limitations of both ML and DL architectures. To provide a more detailed perspective, ML and DL techniques are further examined across applications in ten different disease types. In addition, the role of adipose tissue biomarkers in disease risk prediction is analyzed. Despite these advancements, the review identifies challenges related to dataset limitations, annotation constraints, feature extraction, model interpretability, prediction reliability in disease risk assessment, and integration into clinical practice. Consequently, future research directions should focus on improving imaging modalities, expanding standardized datasets, advancing DL architectures and hybrid models, enhancing generalization and interpretability, and ensuring seamless clinical workflow integration.