Laryngeal Cancer Prediction via VGG16-Based Feature Extraction and Machine Learning on Narrow-Band Imaging Data

Abstract

Laryngeal cancer is a serious type of cancer that poses a significant threat to patients' lives if not detected early, due to its impact on the respiratory system. This study aims to improve an intelligent automated model capable of predicting laryngeal cancer using machine learning and deep learning algorithms. A dataset of narrowband medical images of laryngeal tissues, obtained from the Kaggle website, was obtained and divided into a 60% training set (792 images, 198 images per category), a 20% validation set (264 images, 66 images per category), and a 20% test set (264 images, 66 images per category). The dataset underwent several preprocessing steps, including the application of Gaussian filters to remove noise from the medical images, conversion from RGB to YCBCR for image enhancement using CLAHE, and the application of a sharpening filter to increase the clarity of image details and edges. Data augmentation was also used to diversify and expand the data. The VGG16 model was used to extract deep features from medical images, followed by a performance optimization phase using ANOVA-F to select the best features and a network search to fine-tune hyperparameters, before predicting the outcomes. Subsequently, machine learning algorithms KNN, SVM, and LR, as well as the deep learning model VGG16, were trained on the features extracted from the VGG16 model (256 features per image for the VGG16 classifier, 150 features per image for the KNN and SVM classifiers, and 512 features per image for the LR classifier). Finally, the results showed that the VGG16 model achieved superior performance, with an accuracy of 0.96, a resolution of 0.96, a recall of 0.96, and an F1 score of 0.96, outperforming other machine learning models, where the accuracy of the LR, SVM, and KNN models was 0.93, 0.92, and 0.90, respectively.