建置Lung-RADS肺結節AI輔助診斷系統之研究

Abstract

根據衛福部癌症死亡人數資料顯示，肺癌已蟬聯10年死亡率冠軍，是當前最致命的癌症之一，由於早期篩檢不易，將近7成屬於晚期患者不容易治療，再加上肺癌腫瘤容易轉移的特性，死亡率一直居高不下。近年許多早期篩檢方法被提出，包括使用低劑量電腦斷層 (Low Dose Computed Tomography, LDCT) 取代傳統X光片進行肺癌篩檢，與提出相應的肺部影像報告系統如Lung-RADS (Lung CT Screening Reporting and Data System) 準則以提供標準化的肺癌篩檢診斷流程。然而低劑量電腦斷層影像判讀是相當費時之工作，在可預見的未來肺癌篩檢將持續普及，屆時將有更大量的影像需由醫生判讀，若能引進自動化判讀技術，將可大幅降低判讀人力成本。過往研究中，除了探討判讀流程中的各步驟自動化算法外，也逐漸發展基於肺部影像報告系統之算法整合研究。 本研究提出一基於Lung-RADS準則之肺癌輔助診斷演算流程，旨在大幅節省醫生判讀低劑量電腦斷層影像所需時間。演算流程使用機器學習、影像組學等方法達成自動化。訓練出之模型能直接從電腦斷層影像中偵測肺結節位置並對之進行分割，並可警示類別有爭議之結節，相比過往研究更符合實務診斷邏輯。最後開發一可實務使用之系統，對系統進行基本的可用性測試，將系統演示給相關專業醫師測試後搜集反應意見，提供未來推動產品化方向之參考。

According to the data of the Ministry of Health and Welfare, lung cancer has the highest mortality rate for past 10 years, and is one of the deadliest cancers at present. Since early screening is not easy for lung cancer, nearly 70% of patient belongs to advance stage when they were diagnosed, which is hard to be treated. Furthermore, lung cancer tumors are easy to metastasize, cause the mortality rate remains very high. In recent years, many early screening methods have been proposed, including the use of Low Dose Computed Tomography (LDCT) instead of traditional X-ray films for lung cancer screening, and the corresponding screening reporting system such as Lung-RADS (Lung CT Screening Reporting and Data System) guidelines had been proposed to standardize LDCT lung cancer screening diagnostic procedures. However, the interpretation of LDCT images is quite time-consuming. In the foreseeable future, lung cancer screening will continue to be popularized. By then, a larger number of images will need to be viewed by doctors. If automated screening technology can be introduced, the cost of screening manpower can be greatly reduced. In previous studies, in addition to discussing the automatic algorithm of each step in the screening process, algorithm integration study based on lung imaging reporting system has also been gradually developed. In this study, an automatic lung cancer screening process based on the Lung-RADS guidelines has been proposed, which aims to significantly save the time of doctors. The automatic screening process is automated by using machine learning, radiomics and some other methods. The trained model can directly detect and segment pulmonary nodules from CT images, and can warn the doctors about the nodules with controversial categories, which is more in line with practical diagnosis logic than previous studies. Afterward, a practical system which ensemble the whole process is developed as a software application. At last, basic usability test is performed on the software, and the software was demonstrated to professional doctors and we collected their opinions, providing a reference for future productization.