A Region-Based Convolutional Neural Network Model for Quantitative Analysis of Carbohydrate Content in Foods

Type 1 Diabetes Mellitus (T1DM) is a globally prevalent autoimmune disease, increasing in incidence annually. The condition necessitates continuous monitoring and meticulous record-keeping of dietary intake, with a strong emphasis on accurate carbohydrate counting. Recent advancements in computer vision have facilitated the estimation of nutritional content and values of meals, enabling calculations through both 2D and 3D image analyses. Significant progress in artificial neural networks has further enhanced the accuracy and efficiency of food recognition and volume estimation. In this study, we propose a system that estimates the carbohydrate content of meals by analyzing their diameters as input. The system demonstrated an average error rate of approximately 7%, with individual error rates ranging from 1% to 15%. While these variations are influenced by the presentation style of the food, the system’s ability to achieve high accuracy from a single image highlights its effectiveness. Moreover, the system is adaptable for 3D volume estimation using multi-angle images, making it suitable for further development. With the integration of additional food categories and an expanded training dataset, the proposed system holds significant potential for practical application in dietary management and nutritional monitoring.