Principal Investigator: David W. Bates
Urinalysis has great potential in personalized care. Mobile devices incorporate image sensors, offering a practical, accurate, and low-cost mHealth-based solutions for initial self-diagnosis of disease, self-monitoring of health conditions, or to augment preliminary remote clinical examinations. Our objective is to (1) compare the performance of different algorithms to classify non-uniform illumination of urinalysis strip images for color recognition and (2) evaluate the feasibility of image processing on mobile devices in the context of urinalysis. 5,620 labeled urinalysis strip images were split into a training set (5,163, 91.8%) and hold-out validation set (457, 8.2%). We propose a backpropagation neural network-based algorithm to identify color similarity between urinalysis images originating on mobile devices (under non-uniform illumination) and a standard colorimetric card that will process images on the mobile device itself (novel workflow) or on the server (baseline workflow). Three existing image recognition models, run on the server, serve as additional baselines. Algorithm performance is evaluated using the metrics of average accuracy and improvement rate. There is no evidence that algorithm performance decreases when processing mobile analysis.
Objectives (1) Compare the performance of different algorithms to classify non-uniform illumination of urinalysis strip images for color recognition, and (2) evaluate the feasibility of image processing on mobile devices in the context of urinalysis.
Methods 5,620 labeled urinalysis strip images were split into a training set (5,163, 91.8%) and hold-out validation set (457, 8.2%). We propose a backpropagation neural network-based algorithm to identify color similarity between urinalysis images originating on mobile devices (under non-uniform illumination) and a standard colorimetric card that will process images on the mobile device itself (novel workflow) or on the server (baseline workflow). Three existing image recognition models, run on the server, serve as additional baselines. Algorithm performance is evaluated using the metrics of average accuracy and improvement rate.
Results The result indicates that our algorithm performs stably (average accuracy as 90.9%) and is higher than the four baselines with a maximum improvement rate of 28.2% and an average improvement rate of 16.9%.
Conclusions It is feasible to run image processing algorithms on mobile devices themselves for accurate urinalysis. Mobile urinalysis enables patients to control their own personal data and get vital lab results quickly while enabling them to share data with their care team.
If the PDF viewer does not load initially, please try refreshing the page.
One Response