NOVEL IMAGE-BASED DIETARY ASSESSMENT TOOLS: THE ROLE OF THE MACHINE LEARNING APPROACHES FOR FOOD RECOGNITION AND NUTRITIONAL EVALUATION IN EPIDEMIOLOGICAL STUDIES

Diet is a complex and multidimensional exposure, and quantifying its contribution to chronic disease risk remains a major challenge for nutritional epidemiology. Traditional dietary assessment methods suffer from recall and reporting bias and feasibility constraints in large-scale studies. Recent advances in deep learning (DL) support novel, image-based dietary assessment tools improving accuracy, scalability, and cultural adaptation. This doctoral research contributed to this field through systematic review, cross-country nutrient harmonization, DL-based nutrient profile prediction, and development of web-based tools for epidemiological applications. First, two systematic reviews examined a posteriori dietary patterns (DPs) in Italian populations (2001–2022). The first assessed reproducibility of DPs derived through principal component and factor analysis, identifying 186 DPs consolidated into 11 reproducible groups with moderate-to-high concordance despite heterogeneity in derivation and labeling. The second evaluated whether reproducible DPs were consistently associated with health outcomes, drivers, and correlates. Evidence was heterogeneous and often underpowered; vegetable-based and starchy patterns showed opposing links with gastric cancer. More than half of reported associations were null, highlighting methodological variability and incomplete confounder adjustment. Second, the Nutrition5k dataset was adapted to the Italian context through ingredient-level harmonization with the Italian BDA. Challenges included discrepancies in ingredient definitions, missing values, and portion-size standardization, yet harmonization was feasible and necessary for culturally diverse DL applications. Third, DL models were trained on Nutrition5k images for predicting mass, energy, macronutrients, and recognizing ingredients. Harmonization, manual curation, and automated frame filtering improved performance, though results remained weaker for visually complex dishes. The Vision Transformer achieved the most consistent performance. A pilot dataset of Italian composite recipes with RGB-Depth imaging and full annotation was developed, and a tailored web-based platform with a validated FFQ implemented for the INDACO study. In conclusion, this thesis advances image-based dietary assessment by integrating epidemiologic instruments with DL methods, demonstrating the feasibility of harmonization, the role of data quality in performance, and providing datasets and tools for future applications in epidemiology and public health.