Real-Time Pill Detection and Recognition Framework Based on a Deep Learning Algorithm

Introduction

According to the survey of World Health Organization, after heart disease and cancer, medical error is the third largest cause of mortality, with 250–400 thousand or more deaths per year (Makary & Daniel, 2016). The pandemic of medical error was brought to light by the Research Institute of Medicine, which discovered that the most prevalent avoidable medical error is a pharmaceutical error, which causes over 1.5 million injuries and costs over $3 billion in complication costs alone (James, 2013). Meanwhile, the pharmaceutical sector is advancing day by day and more effective drugs are currently being developed to treat ailments. Due to fast industrial progress, many pharmaceutical enterprises have evolved, producing medications in their distinctive method. According to the report of Agency for Healthcare Research and Quality (2019), the procedure before a patient receives drugs consists of multiple steps, including the clinician’s selection of the appropriate medication and plan, the pharmacist’s interpretation of the prescription, checking drug interactions and allergies and finally, the patient receiving the correct drug in the appropriate quantity.

The healthcare providers should follow the “Five Rights” of pharmaceutical safety, including “administering the appropriate medicine, in the right amount, at the right time, through the right channel, to the right patient.” Medication errors may arise during the dispensing process due to the vast number of tablets available, the closeness of medication labels and their appearance. It might be difficult for the typical individual to distinguish or recognize pills and it necessitates the development of a system that can reliably detect and recognize pills.

Several systems have been developed to classify and identify the pills (Kwon et al., 2022). They are mainly divided into two categories (Raja et al., 2019): manual recognition and automatic recognition. The manual recognition technique needs an operator to enter information about the pill’s features manually. Although the manual input approach is simple, it is less efficient than the automated recognition system. When the database contains many drugs, manual entry becomes time-consuming and necessitates using a human resource. Due to that, the automatic pill identification method is preferable for a massive amount of tablets. Since no manual data entry is required, an automated pill identification system is quick and straightforward to use for a vast number of tablets. The field of automatic pill identification is undergoing a lot of research. The healthcare providers should follow the “Five Rights” of pharmaceutical safety, including “administering the appropriate medicine, in the right amount, at the right time, through the right channel, to the right patient.”

Medication errors may arise during the dispensing process due to the vast number of tablets available, the closeness of medication labels and their appearance. It might be difficult for the typical individual to distinguish or recognize pills and it necessitates the development of a system that can reliably detect and recognize pills. The pill identification challenge was launched in January 2016 for the establishment and implementation to detect pills using images of tablets (Chang et al., 2019). There are 1,000 pill images, 2,000 references images and 5,000 consumer-quality images in the dataset. Pill images are acquired for segregation assessment. The challenge winners created algorithms to detect consumer quality images with average accuracy values of 0.27, 0.09 and 0.08. The existing algorithms recovered only 43%, 25% and 11% of the system’s images from consumer queries. Recovery was performed on 5,000 consumer images. It was a significant accomplishment since pill variants fluctuated every year as new pill types were introduced every day.