Executive Summary
The ability to quickly and reliably be able to identify both prescription and non-prescription medications is not only important for a patient and his caretaker, but also important for pharmacists, nurses, and law enforcement agents. Despite unique FDA-approved identifiers for each pill, distinguishing the differences can be difficult and are often exacerbated by poor lighting, worn edges, and fading imprints. Current pill-recognition software is often resource intensive and performance is unreliable. MSU researchers have addressed these issues with a new technology that offers greatly improved accuracy.
Description of Technology
This MSU technology is a small-footprint pill recognition software intended for mobile phones. It utilizes a novel deep learning model optimization technology. The recognition algorithm is based on a deep Convolutional Neural Network (CNN). Knowledge extracted from three CNNs trained on defining characteristics of pills are used. The optimization technique significantly reduces the footprint of the larger deep learning model with negligible recognition accuracy drop by training a much smaller model that imitates the outputs of the larger model. The smaller size is able to perform low-power, near real-time pill image recognition on smartphone CPUs without cloud offload. The “off-the-grid” capability helps insure user confidentiality as well as allow access for those in rural areas or with poor cellular reception. This technology was awarded First Prize in the 2016 Pill Image Recognition Challenge offered by the U.S. National Library of Medicine (NLM) of the National Institutes of Health (NIH).
Key Benefits
Applications
Patent Status
Patent pending
Licensing Rights Available
Full licensing rights available
Inventors
Mi Zhang, Kai Cao, Xiao Zeng
Tech ID: TEC2017-0027