運用深度學習於定位肺結節電腦斷層影像之探討

肺結節為早期肺癌診斷的重點，目前主要以胸腔電腦斷層(CT)掃描來檢測其健康狀況。由於醫生需要從許多 CT 影像中以人工比對方式來判斷病變的發生與否，因此本研究透過目標檢測方法來訓練深度學習模型，以輔助醫師診斷肺結節於影像中之確切位置，來提升診斷效率和準確率。本研究主要分成五個部分，分別是肺部影像資料集之收集與篩選、肺部影像處理、影像手動圈選、深度學習模型的驗證與測試。本研究在肺部自動化影像處理是先將電腦斷層影像進行二值化、黑白像素轉換、洪水填充法與填補孔洞，接著將影像進行 LabelMe 軟體手動圈選後放入深度學習模型中。在深度學習的部分運用 Mask R-CNN 模型架構進行訓練與辨識，另外，會再將影像放入 YOLO v4 模型架構中訓練並一同進行比較，模型測試指標使用平均類別精確度(mean average precision, mAP)以評估模型辨識影像的成效與準確率，最後再進行影像定位與辨識的結果比較。除了 mAP 數值之外，為了觀測兩種模型的學習效能，皆使用損失函數(Loss Function)與 PR 曲線(Precision-Recall Curve)作為評估模型學習效果的指標。實驗結果顯示，將肺部原始影像與前處理後影像放入 Mask R-CNN 與YOLO v4 兩種模型中進行訓練及測試，其前處理後影像之辨識準確率比原始影像有更高，尤其在 Mask R-CNN 模型之 mAP 達到 65%的結果，經過模型參數調整後，原始影像的 mAP 從 0%提升到 33.67%，且在測試集的肺部影像上皆能準確的圈選及辨識出肺結節所在位置。在 YOLO v4 模型的部分與 Mask R-CNN 具有同樣的結果，前處理影像之準確率比原始影像更高；在 YOLO v4 模型中加入五折交叉驗證的方式後，525 張原始與前處理影像之結果遠比 478 張原始影像的結果來的好，mAP 從原始的 39.8%達到 63.32%，前處理影像的 mAP 達到 66.43%。在兩種模型接加入損失函數與 PR 曲線的觀測指標後，結果表示損失函數圖的曲線趨勢皆會隨著迭帶次數增加從左邊的最高點急速下降，直到損失函數小於 5 之後趨於平緩；PR 曲線圖的趨勢皆會隨著精確度增加從左邊最高點緩慢下降，直到召回率小於 0.8 後急速下降且最後等於 0。本研究結合 Python 程式語言、OpenCV 套件與運用深度學習模型架構等技術，建構一套自動辨識肺部影像的流程，未來可進一步應用於不同的醫學工程應用領域。

關鍵字

肺結節；肺部電腦斷層影像；自動化影像處理；深度學習

並列摘要

Pulmonary nodules are the focus of early lung cancer diagnosis, and chest computed tomography(CT) scans are currently used to detect their health status. Since doctors need to detect the occurrence of lesions by manual comparison from many CT images, this study used the object detection method to train deep learning models to assist doctors in diagnosing the exact location of pulmonary nodules in the images, and to improve the diagnostic efficiency and accuracy. This research is mainly divided into five parts, namely collection and screening of lung image dataset, lung image processing, manual image circle selection, testing and validation of deep learning model. In this study, the first automatic lung image processing is to perform binarization, black and white pixel conversion, flood filling and hole filling of the CT images, and then the images are manually selected by software LabelMe and put into the deep learning model. In the deep learning part, the Mask R-CNN model architecture is used for training and identification. In addition, the images also will be put into the YOLO v4 model architecture for training and comparing with Mask R-CNN. The model test indicator used mAP(mean average precision) to evaluate the model identification and accuracy. Finally, the image recognition results are compared and analyzed. The experimental results show that when original lung images and preprocessing images are put into Mask R-CNN and YOLO v4 models for training and testing, the recognition accuracy of preprocessing images is higher than original images. Especially, the mAP of Mask R-CNN model reached 65%. After adjusting the model parameters and adding the Early-Stopping method, original images mAP increased from 0% to 33.67%. The lung images in the test set can be accurately circled. Select and identify the location of pulmonary nodules. The YOLO v4 model had the same results as Mask R-CNN, and the accuracy of the preprocessed images is higher than original images. After adding the 5-fold cross validation method to the YOLO v4 model, the 525 original and preprocessed images are compared. The results are better than the 478 original images results, its mAP from 39.8% to 63.32%, and preprocessed images mAP reached 66.43%. After adding loss function and PR curve to the two models, the results showed that the trend of the loss function graph will decrease rapidly from the highest point on the left, the increase of the Epoch until loss function value is less than 5. The trend of the PR curve graph will slowly decrease from the highest point on the left as the precision increased, until recall value is less than 0.8 and then decreased rapidly and finally equalled to 0. This study used the Python programming language, OpenCV library, and deep learning model architecture to construct a process for automatically identifying lung images. It can be further applied to different medical engineering application fields in the future.