Part-time Lecturer, Department of Oral and Maxillofacial Radiology, Aichi-Gakuin University School of Dentistry, Nagoya, Japan.
Associate Professor, Department of Oral and Maxillofacial Radiology, Aichi-Gakuin University School of Dentistry, Nagoya, Japan.
Oral Surg Oral Med Oral Pathol Oral Radiol. 2020 Oct;130(4):464-469. doi: 10.1016/j.oooo.2020.04.813. Epub 2020 Jun 2.
This investigation aimed to verify and compare the performance of 3 deep learning systems for classifying maxillary impacted supernumerary teeth (ISTs) in patients with fully erupted incisors.
In total, the study included 550 panoramic radiographs obtained from 275 patients with at least 1 IST and 275 patients without ISTs in the maxillary incisor region. Three learning models were created by using AlexNet, VGG-16, and DetectNet. Four hundred images were randomly selected as training data, and 100 images were assigned as validating and testing data. The remaining 50 images were used as new testing data. The sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve were calculated. Detection performance was evaluated by using recall, precision, and F-measure.
DetectNet generally produced the highest values of diagnostic efficacy. VGG-16 yielded significantly lower values compared with DetectNet and AlexNet. Assessment of the detection performance of DetectNet showed that recall, precision, and F-measure for detection in the incisor region were all 1.0, indicating perfect detection.
DetectNet and AlexNet appear to have potential use in classifying the presence of ISTs in the maxillary incisor region on panoramic radiographs. Additionally, DetectNet would be suitable for automatic detection of this abnormality.
本研究旨在验证和比较 3 种深度学习系统在分类完全萌出切牙区上颌埋伏多生牙（ISTs）方面的性能。
共纳入 275 例至少有 1 颗 IST 和 275 例上颌切牙区无 IST 的患者的 550 张全景片。使用 AlexNet、VGG-16 和 DetectNet 建立了 3 个学习模型。400 张图像被随机选择作为训练数据，100 张图像被分配作为验证和测试数据。其余 50 张图像被用作新的测试数据。计算了灵敏度、特异性、准确性和接收者操作特征曲线下的面积。通过召回率、精度和 F 值来评估检测性能。
DetectNet 通常产生最高的诊断效能值。与 DetectNet 和 AlexNet 相比，VGG-16 的值显著降低。对 DetectNet 的检测性能进行评估，发现其在切牙区的检测召回率、精度和 F 值均为 1.0，表明完美检测。
DetectNet 和 AlexNet 似乎有潜力用于在全景片上分类上颌切牙区 ISTs 的存在。此外，DetectNet 适合于这种异常的自动检测。
