Identifying Medically-compromised Patients with Periodontitis-Associated Cardiovascular Diseases Using Convolutional Neural Network-facilitated Multilabel Classification of Panoramic Radiographs

The bidirectional relationship between periodontitis and atherosclerotic cardiovascular disease (ASCVD) has been demonstrated in cohort studies. In this study, we applied computer vision (CV)-based algorithms and convolutional neural networks (CNNs) to identify periodontitis-associated ASCVD through panoramic radiographs. 432 radiographs were balancedly collected at a medical center, from patients with both ASCVD and periodontitis, with only periodontitis, with only ASCVD, and without either ASCVD or periodontitis. The panoramic radiographs were first segmented with U-Net as original images without any segmentation, images with only the maxilla, images without teeth, images with only the mandible, and images with only teeth. Then, CV-based algorithms for average brightness histogram analysis and CNN-based multi-label classification were parallelly used to recognize two labels, ASCVD and periodontitis. The multi-label classification task was executed with hyperparemeters including adam and binary cross-entropy. Compared to average brightness analysis, the accuracy of multi-label classification for the two labels was satisfying, with the F2 score and recall being 0.90 and 0.93 for original images, respectively. In conclusion, multi-label classification incorporating CNN could better recognize not only periodontitis but ASCVD. Moreover, maxilla played a key role in providing information for classification, which was in line with domain knowledge regarding how ASCVD may involve the head and neck area.