CNN-Based Pill Image Recognition for Retrieval Systems

• Published in: Applied sciences Vol. 13; no. 8; p. 5050
• Main Authors: Al-Hussaeni, Khalil, Karamitsos, Ioannis, Adewumi, Ezekiel, Amawi, Rema M.
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01-04-2023
• Subjects: Accuracy; Analysis; Automation; Caregivers; CBIR; Cellular telephones; Classification; CNN; convolutional neural networks; Datasets; Deep learning; Digital imaging; Image classification; Image enhancement; Image processing; image recognition; Image retrieval; Information processing; Information retrieval; Information storage and retrieval; machine learning; Medical errors; Medicine; Metadata; Neural networks; pill information retrieval; Prescription drugs; Web applications; United Arab Emirates; United States > US
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app13085050

Medication should be consumed as prescribed with little to zero margins for errors, otherwise consequences could be fatal. Due to the pervasiveness of camera-equipped mobile devices, patients and practitioners can easily take photos of unidentified pills to avert erroneous prescriptions or consumption. This area of research goes under the umbrella of information retrieval and, more specifically, image retrieval or recognition. Several studies have been conducted in the area of image retrieval in order to propose accurate models, i.e., accurately matching an input image with stored ones. Recently, neural networks have been shown to be effective in identifying digital images. This study aims to provide an enhancement to image retrieval in terms of accuracy and efficiency through image segmentation and classification. This paper suggests three neural network (CNN) architectures: two models that are hybrid networks paired with a classification method (CNN+SVM and CNN+kNN) and one ResNet-50 network. We perform various preprocessing steps by using several detection techniques on the selected dataset. We conduct extensive experiments using a real-life dataset obtained from the National Library of Medicine database. The results demonstrate that our proposed model is capable of deriving an accuracy of 90.8%. We also provide a comparison of the above-mentioned three models with some existing methods, and we notice that our proposed CNN+kNN architecture improved the pill image retrieval accuracy by 10% compared to existing models.