AUTHOR=Chotwanvirat Phawinpon , Hnoohom Narit , Rojroongwasinkul Nipa , Kriengsinyos Wantanee 

TITLE=Feasibility Study of an Automated Carbohydrate Estimation System Using Thai Food Images in Comparison With Estimation by Dietitians

JOURNAL=Frontiers in Nutrition

VOLUME=Volume 8 - 2021

YEAR=2021

URL=https://www.frontiersin.org/journals/nutrition/articles/10.3389/fnut.2021.732449

DOI=10.3389/fnut.2021.732449

ISSN=2296-861X

ABSTRACT=Carbohydrate counting is essential for well-controlled blood glucose in people with T1D, but to perform it precisely is challenging especially with Thai food. We aimed to develop the Deep Learning-based system for performing carb counting automatically using Thai foods images taken from smartphones. We constructed the new Thai food image dataset was containing 256,178 ingredient objects with measured weight from 175 food categories in 75,232 images. It was used to train both object detector and weight estimator algorithms. After training, the system acquired Top-1 accuracy 80.9% and Root Mean Square Error (RMSE) of carbohydrate estimation less than 10 g in the test dataset. Another set of twenty images, which totally contained 48 food items, were used to compare accuracy of carbohydrate estimations between measured weight, system estimation and eight experienced registered dietitians (RDs). The error of system estimation was 4% while the error of estimations from nearest, lowest, and highest carbohydrate among RDs were 0.7%, 25.5% and 7.6% respectively. The RMSE of carbohydrate estimation of the system and lowest RD were 9.4 and 10.2. The system could perform with estimation error less than 10 g for 13/20 images, it was the third place behind only two of the best performing: RD1 (15/20 images) and RD5 (14/20 images). The system could achieve our satisfaction for estimation of carbohydrate content accuracy and the results were comparable with the experienced dietitians.