| Sumario: | This study aimed to compare the changing trends of gray and texture values of laryngoscopic images in patients with laryngopharyngeal reflux (LPR) and non-LPR. A total of 3428 laryngoscopic images were selected and divided into two groups, non-LPR and LPR groups based on the reflux symptom index. Gray histogram and gray-level co-occurrence matrix (GLCM) were used to quantify gray and texture features, and the model was trained based on these features. The total laryngoscopic images dataset was proportionally split into two parts including the training set and the test set according to the ratio of 7:3. Four different machine learning algorithms, including decision tree, naive Bayes, linear regression, and K-nearest neighbors, were applied to classify non-LPR or LPR laryngoscopic images. The results showed that different classification algorithms are used to classify laryngoscopic image dataset and promising classification accuracy are obtained. Specifically, the accuracy of K-nearest neighbors was 83.38% for the gray histogram-only classification, that of linear regression was 88.63% for the GLCM-only classification, and that of the decision tree was 98.01% for the combined gray histogram and GLCM analysis. Gray histogram and GLCM analysis of the laryngoscopic images may be used as auxiliary tools to detect laryngopharyngeal mucosal damage in patients with LPR. Measurement of gray and texture feature values is an objective and convenient method, which may serve as a reference baseline for clinicians and have potential clinical usefulness.
|
|---|