運用動態模型探討卷積神經網路辨識CT影像內肺部結節之超參數設定研究

Using Dynamic Method to Determine the Hyper-parameters in Convolutional Neural Network for the Classification of Lung Nodules in CT Images

陳祥輝(Hsiang-Hui Chen) ； 黃怡婷(Yi-Ting Hwang) ； 滕涵菁(Harn-Jing Terng)

《中國統計學報》 58卷2期 (2020/06) Pp. 93-127

摘要

肺癌在台灣與工業化國家都是癌症死亡率排名第一位，胸部X光檢查無法滿足肺癌篩檢要求，除了肺部結節可能會被胸骨或其他器官遮蓋之外，X光檢查時僅能提供二維（two-dimension）資訊，對於肺部異常結節（lung nodule）篩檢的敏感度不足。近年來發展3D立體影像技術，運用斷層掃描將切片（slice）影像組合成爲一個三維（3D）立體影像圖形，再經旋轉進一步檢視不同角度，可得到較完整的影像資訊，藉以提升篩檢的敏感度。本研究使用肺圖像數據庫聯盟圖像採集（The Lung Image Database Consortium imagecollection；簡稱LIDC-IDRI）置放在『The Cancer Image Archive（TCIA）Public Access』網站所提供電腦斷層掃描辨識肺部結節的國際標準影像格式資料（Digital Imaging and Communications in Medicine；簡稱；DICOM），該資料包括肺癌診斷和胸部斷層攝影掃描（Computed Tomography；簡稱CT）篩檢，及標示病灶的註釋。本研究透過前置處理DICOM影像，建構卷積神經網路（convolutional neural network，簡稱CNN），瞭解CNN參數設定，進而對肺部結節的進行辨識。由於參數非常多，本論文提出利用動態建模方式來設定這些參數值，最後本研究有找出使用CNN來進行肺部結節辨識的最佳參數值組合。

關鍵字

人工智慧 ； 卷積神經網路 ； 深度學習 ； 動態建模 ； 肺部結核

並列摘要

2D X-rays is often used to screen the lung nodule and a professional physician has to interpret the result. However, the lung nodule can be hidden by thoracic bone and the diagnostic result can be affected. Since there are new inventions in 3D imaging techniques, the low dose computed temography (LDCT) is used to obtain image slices which are then combined into a 3D image. Using the rotation techniques, the image can be viewed from various angles. It expects to increase the accuracy of diagnostic results. This research uses data provided by the Lung Image Database Consortium image collection. The data include lung cancer diagnoses, labelling of specific nodules and computed tomography (CT) scans data. The CT scan image data is stored in a DICOM format. The objective of this study is to clarify the procedure of data pre-processing of DICOM image data for implementing the convolutional neural network (CNN) and then to understand the feasible parameter settings for CNN to classify the lung nodule. An open source software library, TensorFlow is used to perform CNN. Since many parameters have to be specified, a dynamic modelling is proposed to choose the parameter for training the data. Finally, a best parameter setting for classifying the lung nodule for this data is recommended.

並列關鍵字

Artificial intelligence ； Convolutional neural network (CNN) ； Deep learning ； Dynamic modelling ； Lung nodule

參考文獻

Aberle, D. R., Adams, A. M., Berg, C. D. and et al. (2011). Reduced lung-cancer mortality with low-dose computed tomographic screening. New England Journal of Medicine, 365(5), pages 395-409.

Chen, H. (2019). DICOM Processing and Segmentation in Python. https://www.raddq.com/dicom-processing-segmentation-visualization-in-python.

Chian, C. F., Hwang, Y. T., Terng, H. J., Lee, S. C., Chao, T. Y., Chang, H., Ho, C. L., Wu, Y. Y. and Perng, W. C. (2016). Panels of tumor-derived RNA markers in peripheral blood of patients with non-small cell lung cancer: their dependence on age, gender and clinical stage. Oncotarget, 7(31), pages 50582-50595.

Chollet, F. et al., (2015). Keras. https://keras.io.

Chudgar, N. P., Bucciarelli, P. R., Jeffries, E. M., Rizk, N. P., Park, B. J., Adusumilli, P. S. and Jones, D. R. (2015). Results of the national lung cancer screening trail: where are we now? ThoracSurgClin, 25(2), pages 145-153.