透過人工智慧模型判讀牙科環口影像

Abstract

近年來智慧醫療領域發展非常迅速，包含醫學影像的電腦輔助分析，並且也擴展到各種的牙科影像，其中牙科環口影像，或稱為全口X光，是包含口腔區域軟硬組織的一張X光片圖像。在醫院或牙科診所中，它經常被用作評估患者牙齒狀況和促進溝通的前線工具，使牙醫能夠迅速了解患者的牙齒健康狀況。它有助於早期檢測口腔疾病，從而導致更好的預後，並幫助診斷和規劃治療。此外，全口X光還能夠幫助牙醫在治療之前更好地與患者溝通，解釋治療方案和預期結果。本研究主要使用人工智慧模型來自動判讀牙科環口影像，幫助牙醫師在日常臨床工作中有更高的效率。

雖然目前有關牙科環口影像結合人工智慧的研究正蓬勃發展，但大多數的研究目標都集中在單一問題上，例如偵測牙齒、牙齒編號或特定單一發現。有些研究試圖判讀整個影像，但未能有效地融合牙科知識，因此難以正確結合牙齒編號和影像中的牙科相關發現。此外，儘管有幾項研究與牙醫師進行了比較，但也是專注在單一的項目上，並未解決像此研究這樣複雜的問題。最後，大多數研究使用的是相對較小的單一來源資料。

本研究中的人工智慧系統除了偵測牙齒以及其對應之編號外，也檢測臨床發現，如植體、殘根、牙冠/牙橋、根管充填物、補牙、齲齒和根尖囊腫。並使用來自多個國家以及機構的牙科全口X光資料，使模型能夠達到高精度並融入牙科知識。結果顯示，我們的模型經過訓練和測試，能夠在不同的資料集中都準確地識別和分類各種牙科病變及狀況，從而提供詳細的診斷信息。除此之外，透過比較我們的深度學習模型與執業牙醫師的表現，模型的表現與執業牙醫不相上下，並在某些項目中優於牙醫師。

此人工智慧系統可以在日常臨床工作中協助牙醫，減少診斷時間並提高診斷準確性。此外，該模型還具有擴展潛力，可以加入更多牙科臨床發現項目在影像上，或甚至應用於其他醫學影像，進一步提升醫療診斷和治療的效率和效果。總體來說，我們的研究展示了人工智慧在牙科影像分析中的巨大潛力，並為未來的臨床應用提供了堅實的基礎。

Computer-assisted analysis of medical images, particularly through deep learning and artificial intelligence, has rapidly advanced in recent years. These methods have expanded to dental imaging, including dental panoramic images, periapical films, and cone-beam computed tomography, making deep learning a promising approach for dental image analysis.

A dental panoramic radiograph, or orthopantomogram, is an X-ray scan of the oral area. In dental clinics, it is often a frontline tool for assessing patients' dental status and facilitating communication. It helps detect oral diseases early, leading to better prognoses, and aids in diagnosing and planning treatments.

While there has been booming research involving dental panoramic radiographs and artificial intelligence, most of them focus on single issues like tooth detection, numbering, or specific findings. Some attempts to summarize the entire image failed to incorporate dental knowledge effectively, struggling to combine the correct tooth numbers with findings. Additionally, while several studies compared results with human experts, they did not address complex problems like ours. Lastly, many of these studies used relatively small, single-source datasets.

In this work, we provide an end-to-end solution for summarizing dental panoramic images. This involves identifying all teeth and their indices and detecting clinical findings such as tooth number, implants, residual roots, crowns/bridges, root canal fillings, fillings, caries, and periapical radiolucency. Our dataset, sourced from multiple institutions across different countries and ethnicities, allowed our model to achieve high accuracy and incorporate dental knowledge. Comparing our deep learning model's performance with practicing dentists showed comparable generalizability and accuracy. The AI system we developed can assist and pre-warn dentists in their daily clinical work and has the potential to be expanded to include more findings on dental panoramic radiographs or other medical images in the future.