Diabetologie und Stoffwechsel 2015; 10 - P105
DOI: 10.1055/s-0035-1549611

A computer vision-based Smartphone system for carbohydrate counting

S Mougiakakakou 1, 2, M Anthimopoulos 1, J Dehais 1, 3, S Shevchik 1, R Botwey 1, D Duke 4, A Greenburg 4, D Rhyner 1, 2, H Loher 2, J Weissmann 5, C Stettler 2, P Diem 2
  • 1ARTORG Center/University of Bern, Bern, Switzerland
  • 2Bern University Hospital 'Inselspital', Bern, Switzerland
  • 3Graduate School for Cellular and Biomedical Sciences/University of Bern, Bern, Switzerland
  • 4Roche Diagnostics Operations/Diabetes Care, Indianapolis, United States
  • 5Roche Diagnostics Deutschland GmbH, Mannheim, Germany

Introduction: Estimating meals' carbohydrates (CHOs) is important in diabetes. To this end, a computer vision-based smartphone system, named GoCARB, was developed and evaluated. The system is designed to estimate the CHO with error less than ± 20 grams/meal.

Methods: The user places a reference object next to the meal and acquires two images using a smartphone's camera. Then, the different food items on the plate are segmented and recognized while their 3D shape is reconstructed. Based on the shape, the segmentation results and the reference object, the volume of each item is estimated. Finally, the CHO content is calculated by combining the food type with its volume, and using nutritional databases.

Results: GoCARB was evaluated using real meals in laboratory setup and in clinical environment involving 19 adult individuals with type 1 diabetes (T1D). In the first case, GoCARB was able to estimate the CHO content of 27 meals with a mean absolute error of 6 ± 8 CHO grams. In the preclinical study, each participant was asked to count the CHO content of each meal. Then, he/she was asked to estimate the CHO by using GoCARB. A total of 114 meals were used. The mean absolute error was 28.03 ± 38.41 CHO grams of CHO for individuals with T1D and 13.28 ± 10.37 CHO grams by using the GoCARB system.

Conclusions: The evaluation results indicate that GoCARB meets the requirements, while it seems to be more accurate than the average individuals with T1D. Its effectiveness in improving glycemic control will be investigated in a clinical trial.