AIDen: An AI-empowered detection and diagnosis system for jaw lesions using CBCT

牙科CBCT是一种被广泛采用的3D成像方式，用于帮助牙医检测和诊断颌骨病变。由于 最小的信息损失(与传统的2D射线照相相比)和较低的辐射曝光(与 传统的CT)，它已成为各种牙科领域的首选放射学技术。间隙：伴随 牙科CBCT的明显好处是向临床医生提供了压倒性的3D数据。以临床医生为基础 CBCT解释存在观察者间/观察者内一致性低和准确性低的问题。人工智能/深度学习(DL) 有望实现CBCT图像分析自动化，并提供客观、准确的检测和诊断 支持临床决策的能力。然而，由于独特和重要的原因，所做的研究有限。 挑战：(1)牙科CBCT提供由不同口腔的复杂混合组成的3D图像 结构/内容，防止直接使用现有的通用钱包DL算法进行图像分割 并呼吁进行新的DL设计。(2)众所周知，AI/DL是数据饥渴的。很难获得大量的数据 由于复杂的口腔解剖和不可避免的人为错误，准确注释的CBCT图像用于训练DL， 这就需要有效的策略来减少数字图书馆训练的注解工作量。(3)由于这些挑战， 目前用于辅助临床医生进行牙科CBCT判读的软件系统不提供高级AI- 基于病变检测和诊断能力，这使得这个STTR项目及时和重要。我们 最近开发了一种将独特的口腔解剖学融入到DL设计中的DL算法，即 “解剖学受限的致密原位放射治疗(AC-UNET)”。除了提高精确度，AC-UNET还 注释效率高，因为它不仅使用CBCT图像进行训练，而且受到解剖领域的约束 通过新颖的数学编码和基于后验正则化的优化获得知识。应用于 显示有根尖周炎(AP)的CBCT初步数据集，AC-UNET CBCT图像的高精度分割和病变检测，性能优于最先进的DL 算法。我们的长期目标是开发有史以来第一个基于人工智能的软件系统，名为“艾登”，以执行 基于牙科CBCT的各种颌骨的自动分割、病变检测和鉴别诊断 具有高准确性、可靠性和重复性的病变/疾病。艾登将协助临床医生提供 为每个患者提供最佳的治疗方案。我们的第一阶段目标是开发和测试艾登为 病变检测和鉴别诊断侧重于AP，一种高度流行的颌骨病变/疾病。三个目标 (1)优化设计：开发AC-UNET的延伸，以整合更广泛的不同类型的 口腔解剖学知识融入DL设计；(2)优化培训：制定主动学习策略 进一步提高AC-UNET训练的注解效率；(3)临床验证和初步评估 临床决策支持的诊断能力。所有目标将为第二阶段奠定基础，届时端到端 Aiden系统将使用多站点数据集进行构建和验证，并解决各种颌骨病变/疾病。

Dental CBCT is a 3D imaging modality widely adopted to help dentists detect and diagnose jaw lesions. Due to minimum information loss (compared to conventional 2D radiography) and low radiation exposure (compared to conventional CT), it has become the “go-to” radiographic technique in various dental fields. Gaps: Accompanying the clear benefits of dental CBCT is an overwhelming amount of 3D data presented to clinicians. Clinician-based CBCT interpretation suffers from low inter-/intra-observer agreement and low accuracy. AI/Deep Learning (DL) holds great promise to automate CBCT image analysis and provide objective, accurate detection and diagnosis capabilities to support clinical decision. However, limited research has been done due to unique and significant challenges: (1) Dental CBCT provides 3D images composed of a complicated mix of different oral structures/contents, preventing the direct use of existing general-purse DL algorithms for image segmentation and calling for new DL designs. (2) AI/DL is known to be data-hungry. It is very difficult to obtain a large number of accurately-annotated CBCT images to train DL due to complex oral anatomy and inevitable human errors, which calls for efficient strategies to reduce annotation effort for DL training. (3) Due to these challenges, the current software systems used to assist clinicians in dental CBCT interpretation do not provide advanced AI- based lesion detection and diagnosis capabilities, which makes this STTR project timely and important. We recently developed a DL algorithm that integrates unique oral anatomy into the DL design, namely “Anatomically-Constrained dense UNet (AC-UNet)”. In addition to improving accuracy, AC-UNet is also annotation-efficient as it is not only trained using CBCT images but also constrained by anatomical domain knowledge through novel mathematical encoding and posterior regularization-based optimization. Applied to a preliminary dataset of CBCTs with periapical lesions indicative of Apical Periodontitis (AP), AC-UNet achieved high accuracy in segmentation and lesion detection on CBCT images and outperformed state-of-the-art DL algorithms. Our long-term goal is to develop the first-ever AI-based software system called “AIDen” to perform automatic segmentation, lesion detection, and differential diagnosis based on dental CBCT for a variety of jaw lesions/diseases with high accuracy, reliability, and reproducibility. AIDen will assist clinicians in providing optimal treatment decision for each patient. Our Phase-I goal is to develop and test the feasibility of AIDen for lesion detection and differential diagnosis focusing on AP, a highly-prevalent jaw lesion/disease. Three aims are: (1) Optimize design: to develop an extension of AC-UNet to integrate a broader range of different types of oral-anatomical knowledge into the DL design; (2) Optimize training: to develop an Active Learning strategy to further improve annotation efficiency of AC-UNet training; (3) Clinical validation and preliminary assessment of diagnosis capability for clinical decision support. All aims will lay groundwork for Phase-II when an end-to-end AIDen system will be built and validated using multi-site datasets and address a variety of jaw lesions/diseases.