Sasaki Yuki, Sato Koryu, Kobayashi Satomi, Asakura Keiko
Link & Communication Inc, Tokyo, Japan.
Department of Social Epidemiology, Graduate School of Medicine and School of Public Health, Kyoto University, Kyoto, Japan.
JMIR Form Res. 2022 Jan 10;6(1):e31875. doi: 10.2196/31875.
A smartphone image recognition app is expected to be a novel tool for measuring nutrients and food intake, but its performance has not been well evaluated.
We assessed the accuracy of the performance of an image recognition app called CALO mama in terms of the nutrient and food group contents automatically estimated by the app.
We prepared 120 meal samples for which the nutrients and food groups were calculated. Next, we predicted the nutrients and food groups included in the meals from their photographs by using (1) automated image recognition only and (2) manual modification after automatic identification.
Predictions generated using only image recognition were similar to the actual data on the weight of meals and were accurate for 11 out of 30 nutrients and 4 out of 15 food groups. The app underestimated energy, 19 nutrients, and 9 food groups, while it overestimated dairy products and confectioneries. After manual modification, the predictions were similar for energy, accurately capturing the nutrients for 29 out of 30 of meals and the food groups for 10 out of 15 meals. The app underestimated pulses, fruits, and meats, while it overestimated weight, vitamin C, vegetables, and confectioneries.
The results of this study suggest that manual modification after prediction using image recognition improves the performance of the app in assessing the nutrients and food groups of meals. Our findings suggest that image recognition has the potential to achieve a description of the dietary intakes of populations by using "precision nutrition" (a comprehensive and dynamic approach to developing tailored nutritional recommendations) for individuals.
智能手机图像识别应用程序有望成为一种测量营养物质和食物摄入量的新型工具，但其性能尚未得到充分评估。
我们根据一款名为CALO mama的图像识别应用程序自动估算的营养物质和食物类别含量，评估了该应用程序性能的准确性。
我们准备了120份已计算出营养物质和食物类别的膳食样本。接下来，我们通过以下两种方式从膳食照片中预测膳食中包含的营养物质和食物类别：（1）仅使用自动图像识别；（2）自动识别后进行人工修正。
仅使用图像识别生成的预测结果与膳食重量的实际数据相似，对于30种营养物质中的11种以及15种食物类别中的4种预测准确。该应用程序低估了能量、19种营养物质和9种食物类别，同时高估了乳制品和糖果。人工修正后，能量的预测结果相似，准确捕捉到了30份膳食中的29份的营养物质以及15份膳食中的10份的食物类别。该应用程序低估了豆类、水果和肉类，同时高估了重量、维生素C、蔬菜和糖果。
本研究结果表明，使用图像识别进行预测后进行人工修正可提高该应用程序评估膳食营养物质和食物类别的性能。我们的研究结果表明，图像识别有潜力通过对个体采用“精准营养”（一种制定个性化营养建议的全面且动态的方法）来实现对人群饮食摄入量的描述。
