Use of Convolutional Neural Networks in Smartphones for the Identification of Oral Diseases Using a Small Dataset

Image recognition and processing is a suitable tool in systems using machine learning methods. The addition of smartphones as complementary tools in the health area for diagnosis is a fact nowadays due to the advantages they present. Following the trend of providing tools for diagnosis, this researc...

Full description

Saved in:

Bibliographic Details
Main Authors:	Quintero-Rojas, Jormany, González, Jesús David
Format:	Online
Language:	eng spa
Published:	Universidad Pedagógica y Tecnológica de Colombia 2021
Subjects:	dentistry medical technology preventive medicine artificial intelligence machine learning oral diagnosis oral disease identification smartphone odontología tecnología médica medicina preventiva inteligencia artificial aprendizaje automático diagnóstico bucal identificación de enfermedades bucales teléfonos inteligentes
Online Access:	https://revistas.uptc.edu.co/index.php/ingenieria/article/view/11846
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Image recognition and processing is a suitable tool in systems using machine learning methods. The addition of smartphones as complementary tools in the health area for diagnosis is a fact nowadays due to the advantages they present. Following the trend of providing tools for diagnosis, this research aimed to develop a prototype mobile application for the identification of oral lesions, including potentially malignant lesions, based on convolutional neural networks, as early detection of indications of possible types of cancer in the oral cavity. A mobile application was developed for the Android operating system that implemented the TensorFlow library and the Mobilenet V2 convolutional neural network model. The training of the model was performed by transfer learning with a database of 500 images distributed in five classes for recognition (Leukoplakia, Herpes Simplex Virus Type 1, Aphthous stomatitis, Nicotinic stomatitis, and No lesion). The 80% of the images were used for training and 20% for validation. It was obtained that the application presented at least 80% precision in the recognition of four class. The f1-score and area under curve metrics were used to evaluate performance. The developed mobile application presented an acceptable performance with metrics higher than 75% for the recognition of three lesions, on the other hand, it yielded an unfavorable performance lower than 70% for identifying nicotinic stomatitis cases with the chosen dataset.