Classification and Prediction of Smoker Melanosis in Gingiva Using Squeezenet Algorithms Publisher



Srinivasan P¹ ; Sundar S² ; Yadalam PK³ ; Ramadoss R² ; Mosaddad SA⁴ ; Heboyan A^{5, 6} Show Affiliations
Source: Oral Surgery Published:2024

Abstract

Introduction: Gingival darkening, known as smokers' gingival melanosis or smoker's melanosis, is a benign symptom attributed to excessive deposition of melanin pigment. Melanin, responsible for skin, hair, and eye coloration, accumulates excessively, resulting in darkened gums. While not typically a severe health concern, smokers and former smokers may worry about this aesthetic change. This study aims to differentiate and categorize gingival pigmentation in smokers compared to non-smokers using a deep learning-based SqueezeNet. This innovative diagnostic tool holds significant promise in the medical and dental fields for assessing the presence and prognosis of such gingival pigmentations. Methods: The research utilized intraoral images obtained from patients at Saveetha Dental College. Two hundred patients exhibiting oral and gingival pigmentation were enrolled, specifically targeting those with a smoking history and excluding cases of amalgam, metallic, physiological pigmentation, intestinal polyposis, and pigmentation from food allergies. After applying inclusion and exclusion criteria, 102 subjects were selected from the initial pool for further analysis. These subjects were classified into healthy and pigmented gingiva, displaying brownish-black pigmentation. Intraoral photographs were captured and subjected to segmentation and annotation using a SqueezeNet for subsequent analysis. Machine learning techniques were employed for further evaluation, including random forest, logistic Regression, and naive Bayes. Results: The accuracy results indicate the performance of three different machine learning algorithms, namely random forest, logistic regression, and Naive Bayes, on predicting the presence of pigmented gingiva in smokers and non-smokers. Using random forest, logistic regression, and Naive Bayes, respectively, it was discovered that the accuracy of pigmented gingiva in smokers and non-smokers was 88.8% and 92.6%, 93.3% and 95.2%, 95.1%, and 95.6%. A confusion matrix and a Roc graph were produced. Conclusions: The results suggest that machine learning algorithms can be effective in predicting the presence of pigmented gingiva in different populations. The confusion matrix and ROC graph can provide further insights into the model's performance and can be used to optimize the classifier's threshold for prediction. Thus, it allows patients to self-diagnose and doctors to diagnose in their hectic clinical routines. © 2024 British Association of Oral Surgeons and John Wiley & Sons Ltd.