STTR Phase I: A Machine Learning Framework for Comprehensive Dental Caries Detection

这项小型企业技术转让（STTR）第一阶段项目的更广泛/商业影响将是开发一种人工智能软件解决方案，该解决方案能够在数字x射线中自动检测龋齿。常规的蛀牙误诊是一个全球性的挑战；在发达国家，光是龋齿就占到医疗保健费用的5%以上，而牙科保健的重点是修复而不是预防蛀牙。这个项目将为已经被全国20万牙医使用的软件开发一个附加解决方案。该项目产生的技术将使非专业助理能够自动对患者进行分诊、筛查和跟踪，从而为全国和全世界服务不足的社区增加获得口腔护理的机会。这个小型企业技术转移（STTR）第一阶段项目将展示两个关键创新的可行性：(1)使用创新的神经网络算法来检测x射线中的蛀牙的新软件框架，以及(2)世界上最大的牙科x光片数据库，由口腔放射学专家注释。R&amp；D的目标是实现腔检测的高灵敏度和特异性，并确保一致的高质量注释。结果包括：(1)在蛀牙检测方面实现最先进的性能，(2)在检测蛀牙的所有阶段方面优于领域专家，以及(3)使专业人员和非专家都能够使用预测置信度的视觉热图来解释病理。提出的技术具有创新的神经网络结构，用于学习牙科x光片的视觉表示，该结构共同表征数据，同时突出其最显著的属性。使用一个新的和原始的训练程序，该技术将最大限度地利用现有的未标记数据。技术挑战包括在保持最小误报率的同时扩展性能，在各种校准设置下建立互操作性，以及在客户使用的机器类型上以合理的资源成本实现所需的结果水平。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader/commercial impact of this Small Business Technology Transfer (STTR) Phase I project will be the development of an artificial intelligence software solution that enables automated detection of dental cavities in digital X-rays. Routine misdiagnosis of dental cavities (tooth decay) is a global challenge; cavities alone account for over 5% of healthcare costs in developed countries, with dental care focused on repairing rather than preventing tooth decay. This project will develop an add-on solution for software already in use by 200,000 dentists nationally. The technology resulting from this project will allow non-expert assistants to automate the triaging, screening, and tracking of patients, increasing access to oral care for underserved communities nationally and throughout the world.This Small Business Technology Transfer (STTR) Phase I project will demonstrate the feasibility of two key innovations: (1) a novel software framework using an innovative neural network algorithm for the detection of cavities in X-rays, and (2) the world’s largest database of dental radiographs annotated by specialists in oral radiology. The goals of R&D are to achieve high sensitivity and specificity in cavity detection and to ensure consistent high-quality annotations. Outcomes include: (1) achieving state-of-the-art performance in cavity detection, (2) outperforming domain experts in detecting all stages of cavities, and (3) enabling professionals and non-experts alike to interpret pathologies using a visual heatmap of prediction confidence. The proposed technology features an innovative neural network structure for learning visual representations of dental radiographs that jointly characterize the data while highlighting their most salient attributes. Using a new and original training procedure, the technology will maximize the benefit of existing unlabeled data. Technical challenges include scaling performance while maintaining a minimal false-negative rate, establishing interoperability under various calibration settings, and achieving the desired level of results on the types of machines used by customers with reasonable resource costs.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.