Deep Learning Approaches for Synthetic Images of Pharmaceutical Drugs and Vitamin Products Analysis

Abstract

Drug labels and packaging inserts play a critical role throughout the pharmaceutical supply chain, from manufacturing and distribution to final consumption. The images of drug labels (also called display panels) can be used to detect illegal, counterfeit, unapproved or potentially harmful pharmaceutical products. But the manual inspection and verification process can be tedious and cumbersome, which demands for automatic and intelligent identification systems. In this research work, a machine learning-based framework is proposed for image analysis of pharmaceutical drugs and vitamins using the National Library of Medicine (NLM) Pill Image Dataset. Pharmaceutical images were extracted to train, validate and test a Random Forest (RF) classifier to distinguish pharmaceutical products. Accuracy, precision, recall and F1 score were evaluated to assess the proposed model. The experimental results showed that the RF model was able to give an accuracy (ACC) of 98.5%, precision (PRE) of 99.9%, recall (REC) of 98.5% and F1-score (F1) of 98.9%. Moreover, the comparison between the proposed RF model and deep learning models such as ResNet-50, CNN and YOLOv5s showed that the RF model achieved better classification ACC. The results show the success of machine learning approaches to automate drug and vitamin recognition, which is a solution that can be used for medication identification, counterfeit drugs detection, and healthcare decision-support applications.