MEDiDEN: Automatic Medicine Identification Using a Deep Convolutional Neural Network

This paper presents MEDiDEN: a smart phone application for automatic medicine identification and medication reminders. The main features of MEDiDEN include the classification of medicine packages, a reminder function that can be set in details and integrates with the mobile OS's notification system, and news feeds for medication-related articles. The most innovative function of the application is medicine classification, which was implemented using the Inception deep-learning architecture. For medicine package classification, the researchers compared the performance of Inception-V3and Inception- V4 with the data in this work. The two models could identify medicine with 92.75% and 94.85% accuracy, respectively. Even though Inception-V4 provided slightly better results, the researchers selected Inception-V3as the model for deployment due to its smaller size, which has the ability to speed up inference run time. The server consists of a Python backend used to run the neural network model. The client application is available on the Android platform. Actual use testing showed that the application could correctly and consistently identify the medicine in the training data.