Artificial intelligence-driven novel tool for tooth detection and segmentation on panoramic radiographs

Abstract

Objective To evaluate the performance of a new artificial intelligence (AI)-driven tool for tooth detection and segmentation on panoramic radiographs. Materials and methods In total, 153 radiographs were collected. A dentomaxillofacial radiologist labeled and segmented each tooth, serving as the ground truth. Class-agnostic crops with one tooth resulted in 3576 training teeth. The AI-driven tool combined two deep convolutional neural networks with expert refinement. Accuracy of the system to detect and segment teeth was the primary outcome, time analysis secondary. The Kruskal-Wallis test was used to evaluate differences of performance metrics among teeth groups and different devices and chi-square test to verify associations among the amount of corrections, presence of false positive and false negative, and crown and root parts of teeth with potential AI misinterpretations. Results The system achieved a sensitivity of 98.9% and a precision of 99.6% for tooth detection. For segmenting teeth, lower canines presented best results with the following values for intersection over union, precision, recall, F1-score, and Hausdorff distances: 95.3%, 96.9%, 98.3%, 97.5%, and 7.9, respectively. Although still above 90%, segmentation results for both upper and lower molars were somewhat lower. The method showed a clinically significant reduction of 67% of the time consumed for the manual. Conclusions The AI tool yielded a highly accurate and fast performance for detecting and segmenting teeth, faster than the ground truth alone. Clinical significance An innovative clinical AI-driven tool showed a faster and more accurate performance to detect and segment teeth on panoramic radiographs compared with manual segmentation.

Performance of deep learning technology for evaluation of positioning quality in periapical radiography of the maxillary canine

“…This discrepancy can be partially attributed to the differences in root configuration between the teeth, indicating the difference in the number of roots per tooth. However, there have been some reports in which all types of teeth, including the maxillary canines, were segmented on panoramic radiographs [6,7,18]. Leite et al reported good performance at segmenting the maxillary canines (recall, precision, and F measure of 0.969, 0.964, and 0.973, respectively) [7].…”

“…However, there have been some reports in which all types of teeth, including the maxillary canines, were segmented on panoramic radiographs [6,7,18]. Leite et al reported good performance at segmenting the maxillary canines (recall, precision, and F measure of 0.969, 0.964, and 0.973, respectively) [7]. Lee et al also reported high segmentation accuracy of 0.889 for the maxillary canines [6].…”

“…lesions [5], and segmentation of teeth [6,7] and the mental foramen [8]. Regarding periapical radiographs, several applications have been reported, including automatic film mounting [9], teeth detection and numbering [10], and segmentation of teeth and lesions [11].…”

Deep Learning-Based Prediction of Paresthesia after Third Molar Extraction: A Preliminary Study

“…In the field of dentistry, research has been conducted about the detection and classification of anatomical variables [ 9 , 10 ], periapical lesions [ 11 , 12 ], dental caries [ 13 , 14 , 15 ], periodontitis [ 16 ], and benign tumors and cysts [ 17 ]. Deep learning analysis has also been applied to cephalometric images, such as detection of landmarks [ 18 ], prediction of the necessity for orthognathic surgery [ 19 ], and detection of cranio-spinal differences [ 20 ].…”

Impact of Image Context on Deep Learning for Classification of Teeth on Radiographs

“…CNNs and variants thereof are commonly used to detect, segment and classify anatomic structures (hard or soft tissue landmarks, teeth) or pathologies (caries, periodontal bone loss, apical lesions, among others) [1]. For tooth classification, for example, models with sensitivities and specificities around 95-98% have been developed [2,3]. CNN learn from pairs of imagery data and labels (e.g., image labels, boxes encapsulating objects of interest, pixel masks) in a supervised way, eventually establishing a statistically based mapping of the input image to the output label.…”