Squeezenet-based Image Classification and Prediction of High Frenal Attachment in the Upper Lip

Abstract

Background & Aim: Traditional methods like clinical examinations and the Graber test have been used to assess abnormal frenula. Still, machine learning offers a quantitative and data-driven methodology to explore the intricate relationship between frenulum variations and their potential consequences. These attachments are linked to age, gender, and oral hygiene. Midline labial frenulum (MLF) abnormalities are common in preschoolers but diminish with age. AI can improve dentistry by predicting and classifying frenum attachments using machine learning algorithms, computer vision tools, big data analysis, predictive analytics, automation, workflow enhancement, and training. Deep learning algorithms can classify attachment types on large datasets, while computer vision tools can efficiently analyze images for specific features. This study aims to explore the predictive modeling and classification of high frenal attachment in the upper lip using machine learning algorithms.
Material and Methods: We collected 180 radiographic image samples, leveraging an institutional computational database. We employed a robust predictive model using the SqueezeNet architecture for image embeddings and subjected to Naive Bayes and Logistic Regression algorithms. The dataset, containing 90 samples with and 90 without high frenum, underwent preprocessing, customization, and division for training and testing. Performance evaluation involved metrics such as AUC, Classification Accuracy, F1 Score, Precision, and Recall.
Results: The Naive Bayes model demonstrated a remarkable AUC of 0.998, while Logistic Regression exhibited perfection with an AUC of 1.000. Both models showcased high Classification Accuracy (Naive Bayes: 0.971, Logistic Regression: 1.000) and balanced F1 Scores. Precision and Recall metrics reinforced the models' accuracy in positive predictions and capturing positive instances.
Conclusion: Our results signify the potential clinical utility of machine learning algorithms in predicting high frenal attachment. Despite limitations, these findings pave the way for enhanced diagnostic capabilities in orofacial anatomy, emphasizing the need for continued research and refinement of predictive models for broader applications