基於深度學習之肺結核胸部X光電腦輔助診斷系統

A computer-aided system based on deep learning for diagnosing tuberculosis using chest X-ray

肺結核為全球十大死因之一，於2018年，全球估計有1,000萬人患有結核病，且共有150萬人死於結核病(WHO, 2020)。根據台灣結核病防治年報，台灣於2018年肺結核新發個案數為9,179人，發生率為每十萬人口38.9，然而面對如此的傳染性疾病，痰結核菌培養貴為結核病的黃金診斷標準，卻需耗費大量的時間，且目前台灣針對疑似肺結核(CXR TB Positive)以及確診肺結核(TB Positive)進行三分類的深度學習模型稀少，有鑑於此，本研究將提出一電腦輔助診斷系統，解決此一問題。 本研究欲針對肺結核胸部X-ray影像，建構一疑似肺結核及確診肺結核分類的最佳化模型及其權重參數，並探討模型於敏感度100%與陰性預測值100%時，臨床醫師之負載率(Loading Rate)，同時設計一使用者介面(Graphical User Interface, GUI)與深度學習模型界接。本研究應用深度學習技術於桃園市某國軍體系個案醫院及美國國家衛生研究院(NIH)，以直方圖校正(Histogram Equalization)以因應來自不同來源的影像，後同時將陽性樣本過採樣(Oversampling)與陰性樣本下採樣(Downsampling)，並使用遷移學習(Transfer Learning)，將預訓練於ImageNet之神經網路於新的辨識任務中進行訓練，使神經網路權重得以迭代更新，並以拔靴法(Bootstrapping)將上一代拔靴的最佳模型做為下一代拔靴的初始權重，以解決樣本分布不均與過度擬合的問題。本研究使用拔靴法所訓練的深度學習模型，以接受者操作特徵曲線(ROC Curve)切出之閾值為標準，於肺結核辨識任務中準確度最高達99.82%，敏感度達100%，AUC達100%；疑似肺結核任務中，準確度最高達61.21%，敏感度62.50%，AUC達60%。在調整閾值至敏感度100%之後，肺結核辨識任務中負載率最低達0.67%，準確度達99.82%；疑似肺結核任務中，負載率最低達70.25%，準確度達29.79%。為符合臨床篩檢實務的需要，本研究最終所建構的電腦輔助診斷系統，在敏感度100%且陰性預測值100%的條件下，於最終驗證中肺結核辨識任務負載率為0.078%，約可降低醫師閱讀醫學影像負擔99.2%；疑似肺結核辨識任務負載率約84%，約可以降低醫師閱讀醫學影像負擔約16%。傳染病如肺結核等，需特別注意模型之敏感度與陰性預測值，深度學習模型的預測機率切點是一大議題，在高敏感度及高陰性預測值的精神下，電腦輔助診斷準確率的上升，能讓放射科醫師閱讀醫學影像的負載比率下降。

Tuberculosis (TB) is one of the top 10 causes of death worldwide. In 2018, an estimated 10 million people suffered from TB and 1.5 million died. According to the Taiwan Tuberculosis Control Report 2018, the incidence of TB was reported at 38.9 per 100,000 population, with 9,179 new cases. However, in the face of such an infectious disease, culture or molecular methods as the gold standard for diagnosing TB are time-consuming. Furthermore, fewer models classify TB into 3 categories, normal, suspected TB, which in this research means chest X-ray (CXR) TB positive, and TB positive. In view of this, this research will propose a computer-aided diagnosis system to solve this problem. Our research aims to construct an optimised model and its parameters for classifying suspected TB and TB positive chest x-ray images. We will discuss the workload of radiologists under the condition of 100% sensitivity and 100% negative predictive value simultaneously. We also design a graphical user interface (GUI) to interface with the deep learning model.We applied deep learning technology to train chest X-rays from Taoyuan Military General Hospital and the National Institutes of Health. Histogram equalisation is used to adapt images from different sources. The positive samples are oversampled and the negative samples are downsampled at the same time, transfer learning is used. A neural network pre-trained from ImageNet is trained on a new classification task to iteratively update the weights of the neural network. In addition, bootstrapping is used to take the best model from the previous bootstrap as the initial weight of the next training bootstrap to solve the problem of uneven sample distribution and overfitting. Using the threshold of the receiver operating characteristic (ROC) curve, our experiments achieve a high accuracy of 99.82%, sensitivity of 100% and AUC of 100% in the task of TB classification; an accuracy of up to 61.21%, sensitivity of 62.50% and AUC of 60% in the task of suspected TB classification. After adjusting the threshold to the circumstance of 100% sensitivity, a minimum load rate of 0.67%, accuracy of 99.82% in the task of TB classification; a minimum load rate of 70.25%, accuracy of 29.79% in the task of suspected TB classification. In order to meet the needs of clinical screening practice, the loading rate of our computer-assisted system is 0.078% and 84%, respectively, in the final validation when classifying TB and suspected TB, which means that the computer-assisted system can reduce the burden of radiologists reading medical images by 99.2% and 16%, respectively. We should note the threshold of the screening model when faced with an infectious disease such as TB. Under the circumstance of high sensitivity and high negative predictive value, the increase in accuracy of the computer-aided diagnosis system can reduce the burden of radiologists in interpreting medical images.