Identification of Laryngeal Lesions Based on Narrowband Endoscopy Imaging Using Artificial Neural Networks and Visual Programming

Abstract

Certain types of lesions on the laryngeal mucosa may indicate the early stages of laryngeal squamous cell carcinoma (LSCC), which constitutes 98% of malignant laryngeal tumors. This study aims to develop artificial intelligence-based methods to classify laryngeal lesions using digital images obtained through narrowband endoscopy imaging. A total of 1,320 digital images of laryngeal tissue, both healthy and lesioned, were classified into four categories. Five machine learning models were developed, utilizing conventional deep convolutional neural networks (CNNs) and capsule networks (CapsNet): VGG16, VGG19, Inception V3, CapsNet without data augmentation, and CapsNet with data augmentation. The latter used images synthetically generated by an adversarial generative network (GAN). These algorithms were implemented using the Orange visual programming software and the Colab computational platform. The inclusion of GAN-enhanced data augmentation significantly improved the performance of the CapsNet classifier across all lesion types. The CapsNet model with GAN data augmentation achieved an average recall, accuracy, and F1 score of 94.7%, marking it as the second-best performing model. The highest performance was achieved by the CNN Inception V3 model, with 97% recall, accuracy, and F1 score, facilitated through visual programming. The combination of CapsNet with GAN-based data augmentation presents a viable alternative for the classification of medical images. The use of the Orange visual programming tool enabled high classification performance—97% in both accuracy and sensitivity—at low computational costs, without the need for advanced programming skills from the user.