Skip to main content
NUTRITION 2026 Main banner

(PTFS06-01-26) NutriScan: An AI-Powered, Image-Based Research Tool for Low-Burden Dietary Assessment

Sunday, July 26, 2026
11:00 AM - 11:04 AM ET
Location: Prince George's Hall B-E (The Hub) - Poster Theater 2 / Poster Board Number: PT2-1

    Presenting Author(s)

  • Darren Liu, MS

    Graduate Student
    Emory University
    Atlanta, Georgia, United States

    Disclosure(s): No relevant financial relationship(s) with ineligible companies to disclose.

    • Co-Author(s)

  • Hanqi Luo, PhD

    Assistant Research Professor
    Emory University
    Oakland, Georgia, United States

  • XH

    Xiao Hu, PhD

    Emory University, United States

  • Runze Yan, PhD

    Assistant Professor
    Emory University
    Atlanta, Georgia, United States

Disclosure(s):
Darren Liu, MS: No relevant financial relationship(s) with ineligible companies to disclose.
Runze Yan, PhD: No relevant financial relationship(s) with ineligible companies to disclose.
Objectives: 24-hour dietary recalls (24HRs) are a widely used and well-validated dietary assessment tool; however, they place a considerable burden on both participants and interviewers, limiting scalability in large-scale research. We developed and validated NutriScan, an AI-powered, image-based dietary assessment tool designed to complement 24HRs by automating food identification and nutrient estimation.

Methods: NutriScan uses large language models with retrieval-augmented generation to automatically identify foods, assign USDA Food and Nutrient Database for Dietary Studies codes, estimate portion weights from meal photos, and generate nutrient profiles for over 65 macro- and micronutrients without requiring trained dietitians. Its accuracy was evaluated on two datasets: 1) ASA24 (3,000 standardized food images with 3 portion sizes each) was used to assess estimation of food weight and macronutrients (energy, fat, carbohydrates, protein) against three commercial image-based tools (Calorie MAMA, Foodvisor, SnapCalorie) using mean absolute error (MAE), with paired t-tests for significance. 2) The Surveying Nutrient Assessment with Photographs of Meals (SNAPMe) dataset (n = 95 participants, 275 study days, 5,000 food images) provided paired before-and-after meal photographs with concurrent 24HRs as a reference for both macro- and micro-nutrients.

Results: On the ASA24 dataset, NutriScan outperformed all three commercial image-based tools (Calorie MAMA, Foodvisor, and SnapCalorie) across all macronutrients (all p < 0.05), achieving MAEs of 49.4 g for food weight, 47.7 kcal for energy, 1.8 g for fat, 6.95 g for carbohydrates, and 1.38 g for protein — an overall 63% reduction in estimation error. On the SNAPMe dataset, MAEs were 71 g for food weight, 91.4 kcal for energy, 4.97 g for fat, 9.72 g for carbohydrates, and 4.21 g for protein. Across both datasets, micronutrient estimation error ranged from 19% to 33% across all 65+ nutrients.

Conclusions: NutriScan achieves reasonable agreement with traditional 24HRs while substantially reducing burden on participants and registered dietitians, supporting its potential for scalable nutrition research.

Funding Sources: None