Automatic Site-Specific Multiple Level Gum Disease Detection Based On Convolution Neural Network

Gum diseases,including the mild form gingivitis and more severe form periodontitis,is one of the most common dental plaque（bacterial biofilm）-initiated dental diseases.The plaque builds up along the gum margins and causes inflammation.It manifests as increased gingival redness（color）,increased volume（edema）,and loss of superficial characteristics（loss of “stippling” from the gum fiber attachment）.They are site-specific and are usually restricted in the gingival area 3mm from the gingival margin.Dentists can identify them through clinical visual inspection or intraoral photography.It has become clinical practice for regular dental check-ups.The aim of this study was to train a computer to identify sites of gingival inflammation at the pixel level through a deep learning approach.We collected standard intraoral photographs of 110 patients and extracted 337 and 110 images for training and validation.Image annotations at specific gingival sites were classified and validated by dental specialists with more than 15 years of clinical experience.The training images are marked with four health status levels（healthy,questionable healthy,questionable diseased,and diseased）.In pixel-level gingival inflammation detection applications,the main challenge is being able to accurately classify various health states that exhibit very similar image features.In this study,we study two approaches:（1）Coarse inflammation region detection and（2）Pixel level inflammation detection on local regions.（1）Coarse inflammation region detection model: We adopt the YOLOv5 s network model for detecting the target gingival inflammation regions.Current feature extraction modules perform poorly due to the ambiguous image features exhibited by different classes of inflammation.We propose to enhance the feature extraction module by employing attention network modules.We also design a specific loss function to enhance the discrimination of various inflammation levels.The metrics for evaluating this model are recall and mean Average Precision（m AP）.It achieves the recall of 0.679 and the m AP of 0.323.（2）Pixel level inflammation detection model: We propose a deep learning based semantic segmentation architecture with a Deep Labv3 plus network with Xception65 and Mobile Netv2 as the backbone for pixel-level classification and prediction of gingival health in the intraoral images.To fully exploit the unique image features of inflammation,such as redness,swelling and smooth shiny caused by lack of texture,a specific attention module is introduced in the image feature extraction layer and a loss function is designed to enhance the segmentation performance of the model.The metrics for evaluating this model is the mean Intersection over Union（MIo U）.It achieves the MIo U of 0.650.The experimental results demonstrate the effectiveness of the proposed system.Furthermore,a mobile phone APP is also developed to perform inflammation detection.The trained network,based on the standard intraoral photographs,is trimmed down to fit into the mobile platform for inflammation detection.The pictures taken from the mobile APP are also assessed by a dentist.A preliminary validation test,compared to the result by a dentist’s assessment,shows that the mobile APP closely match to that of a dentist’s assessment.This shows the feasibility of using a mobile app for dental self-exams,especially when visiting a dentist is difficult or even impossible such as during the COVID-19 pandemic.