A Two-Level Food Classification System For People With Diabetes Mellitus Using Convolutional Neural Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Nutritional Science

Background:

• Accurate macronutrient estimation is crucial for managing Diabetes Mellitus (DM) and determining insulin dosage.
• Existing methods may not adequately address the specific nutritional needs of individuals with DM.

Purpose of the Study:

• To develop and evaluate a food image classification system tailored for the dietary management of Diabetes Mellitus.
• To improve the accuracy of food macronutrient content estimation for diabetic patients.

Main Methods:

• A two-level classification system using Convolutional Neural Networks (CNNs) was designed.
• A new dataset, NTUA-Food 2017 (3248 images, 82 foods), was created for training and validation.
• A novel evaluation metric was introduced, penalizing errors based on postprandial blood sugar discrepancies.

Main Results:

• The system achieved 84.18% and 85.94% accuracy at the first and second classification levels on the NTUA-Food 2017 dataset.
• The first-level classification algorithm improved accuracy on the Food Image Dataset (FID) to 97.08%.
• The mean error in carbohydrate content estimation was less than 2g per serving on the NTUA-Food 2017 dataset.

Conclusions:

• The proposed food image classification system effectively estimates macronutrient content for individuals with DM.
• This technology has the potential to enhance self-management of Diabetes Mellitus through improved insulin dosing accuracy.
• The developed system and evaluation metric offer advancements in applying AI to personalized nutrition.