Pill Image Classification using Machine Learning

Pill classification and recognition are crucial tasks in preventing the misuse of medication. There is a growing need for automation for these tasks due to its inherent complexity. The Food and Drug Administration (FDA) establishes a unique visual appearance to pills based on their shape, color, texture, and imprint information. We propose an automatic classification system for pill images based on their shape and color. Thus, we use image processing techniques to specify an attribute set used by Support Vector Machines and Multilayer Perceptron classifiers. We carry out experiments on a subset of the NLM PIR dataset, provided by the National Library of Medicine. The results indicate that all classifiers perform accurate predictions, with an average accuracy above 99.3%. This high classification accuracy happens even in the presence of unbalanced classes, with precision and recall average scores above 98%. Ergo, the results corroborate the efficiency of our proposal.

I. Introduction

Medication safety is a critical issue in daily patient care. In recent reports, the U.S. Institute of Medicine indicates that medication error is the most preventable medical error [1]. Errors in the prescription or administration of a drug pill may lead to adverse drug events (ADE) - injuries caused by medication intervention related to a drug [2]. The correct identification of drug pills has the potential to improve treatment from patient experience to poison control centers.