Classification of the Relationship between Mandibular Third Molar and Inferior Alveolar Nerve based on Generated Mask Images

Abstract

In recent dentistry research, deep learning techniques have been employed for various tasks, including detecting and segmenting third molars and inferior alveolar nerves, as well as classifying their positional relationships. Prior studies using convolutional neural networks (CNNs) have successfully detected the adjacent area of the third molar and automatically classified the relationship between the inferior alveolar nerves. However, deep learning models have limitations in learning the diverse patterns of teeth and nerves due to variations in their shape, angle, and size across individuals. Moreover, unlike object classification, relationship classification is influenced by the proximity of teeth and nerves, making it challenging to accurately interpret the classified samples. To address these challenges, we propose a masking image-based classification system. The primary goal of this system is to enhance the classification performance of the relationship between the third molar and inferior alveolar nerve while providing diagnostic evidence to support the classification. Our proposed system operates by detecting the adjacent areas of the third molar, including the inferior alveolar nerve, in panoramic radiographs (PR). Subsequently, it generates masked images of the inferior alveolar nerve and third molar within the extracted regions of interest. Finally, it performs the classification of the relationship between the third molar and inferior alveolar nerve using these masked images. The system achieved a mean average precision (mAP) of 0.885 in detecting the region of interest in the third molar. Furthermore, the performance of the existing CNN-based positional relationship classification was evaluated using four classification models, resulting in an average accuracy of 0.795. For the segmentation task, the third molar and inferior alveolar nerve in the detected region of interest exhibited a dice similarity coefficient (DSC) of 0.961 and 0.820, respectively. Regarding the proposed masking image-based classification, it demonstrated an accuracy of 0.832, outperforming the existing method by approximately 3%, thus confirming the superiority of our proposed system. Author