The principle objective of GoCARB is the design, development and evaluation of a computational system which will support individuals with type 1 diabetes in automatically estimating the grams of carbohydrate in a meal in near real-time. GoCARB introduces the concept of the image analysis of meal pictures captured by a smartphone camera, in order to recognize the food on a plate and to estimate the corresponding carbohydrates. This information can be used to estimate the insulin dose. The system is based on a food database, advanced computer vision methods, machine learning algorithms and smartphone technologies.
The system consists mainly of the following modules.
Plate detection and food segmentation: The plate area is detected and the different food items localized and segmented.
Food recognition: Each segmented food item is automatically recognized based on advanced computer vision and machine learning algorithms.
3D model reconstruction and volume estimation: The 3D shape of the segmented and recognized food items is calculated, in order to estimate the corresponding volume.
Carbohydrate (CHO) estimation: Based on the volume, the USDA database is used for the estimation of the carbohydrate content of the meal.
Insulin bolus calculator: Model-based and heuristic approaches are applied for the optimization of the insulin bolus dose using the carbohydrate information. The additional collection of patient specific parameters permit the personalization of the bolus estimation.