Associate Professor and Chair, Predoctoral Program in Pediatric Dentistry, Department of Developmental Sciences, School of Dentistry, Marquette University, Milwaukee, Wis., USA.
Clinical Assistant Professor and Director, Technological Innovation Center, School of Dentistry, Marquette University, Milwaukee, Wis., USA.
Pediatr Dent. 2024 Sep 15;46(5):332-336.
To develop a no-code artificial intelligence (AI) model capable of identifying primary proximal surface caries using bitewings among pediatric patients. One hundred bitewing radiographs acquired at pediatric dental clinics were anonymized and reviewed. The inclusion criteria encompassed bitewing radiographs of adequate diagnostic quality of primary and mixed-dentition stages. The exclusion criteria included artifacts related to sensors' quality, positioning errors, and motion. Sixty-six bitewing radiographs were selected. Images were uploaded to LandingLens™, a no-code AI platform. A calibrated consensus panel determined the presence or absence of proximal caries lesions on all surfaces (using ground truth labeling). The radiographs were divided into training (70 percent), development (20 percent), and testing (10 percent) subsets. Data augmentation techniques were applied to artificially increase the sample size. Sensitivity, specificity, accuracy, precision, F1-score, and receiver operating characteristic area under the curve (ROC-AUC) were calculated. Among the 755 proximal surfaces identified from 66 bitewings, 178 were annotated as caries lesions by experts. The model achieved the following metrics: 88.8 percent sensitivity, 98.8 percent specificity, 95.8 percent precision, 96.4 percent accuracy, and an F1-score of 92 percent by surface. The ROC-AUC was 0.965. The developed model demonstrated strong performance despite the limited dataset size. This may be attributed to the exclusion of unsuitable radiographs and the use of expert-labeled ground truth annotations. The utilization of no-code artificial intelligence may improve outcomes for computer vision tasks.
开发一种无代码人工智能 (AI) 模型，能够在儿科患者的口内片中识别原发性近表面龋。从儿科牙科诊所获得的 100 张口内片被匿名并进行了审查。纳入标准包括原发性和混合牙列阶段具有足够诊断质量的口内片。排除标准包括与传感器质量、定位错误和运动有关的伪影。选择了 66 张口内片。将图像上传到 LandingLens™，这是一个无代码的 AI 平台。一个经过校准的共识小组根据所有表面（使用真实标签）确定近龋病变的存在或不存在。这些射线照片被分为训练（70%）、开发（20%）和测试（10%）子集。应用数据扩充技术人为地增加样本量。计算了灵敏度、特异性、准确性、精度、F1 分数和接收器操作特性曲线下的面积（ROC-AUC）。在从 66 张口内片中识别的 755 个近表面中，有 178 个被专家标记为龋病病变。该模型实现了以下指标：按表面计算，灵敏度为 88.8%，特异性为 98.8%，精度为 95.8%，准确性为 96.4%，F1 得分为 92%。ROC-AUC 为 0.965。尽管数据集规模有限，但开发的模型仍表现出强大的性能。这可能归因于排除了不合适的射线照片和使用了专家标记的真实标签注释。无代码人工智能的利用可能会改善计算机视觉任务的结果。
