基於密集神經網路之肺部電腦斷層掃瞄輔助診斷

Lung CT Computer-Aided Diagnosis in Densely Connected Network

洪靖凱(Jing-Kai Hong)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201803038

摘要

肺癌是現今最常見的死亡原因之一，而對於癌症，及早的診斷及治療才是最有效的治癒方式。然而，對於診斷每個病患都必須檢視過數百張的電腦斷層掃描影像是十分費時的。近年來，卷積神經網路已被應用於醫學影像領域中的病變偵測、切割及分類，而密集神經網路是其中一著名的卷積神經網路架構。因此，本次研究提出一個基於多尺度密集神經網路的電腦輔助診斷系統用來診斷結節的良惡性，希望能有效醫師的負擔。而所提出的多尺度密集神經網路與密集神經網路最主要的差別在於，取代單一輸入的設計，對於一個結節取用了三種尺寸的影像作為輸入，並大量的減少超參數。此外，為了提高性能，模型的訓練及測試期間，使用了兩個開源數據集進行包含自我訓練的預訓練，及利用選擇性採樣方法在我們的數據集上微調。在此次研究中，我們的數據集共有138筆良性及70筆惡性樣本。在此次實驗中，所提出的電腦輔助診斷系統達到出色的性能，準確性93.7%、敏感性93.6%及專一性93.6%與AUC值0.973。

關鍵字

肺腺癌 ； 電腦斷層掃描 ； 電腦輔助診斷 ； 密集神經網路 ； 自我訓練 ； 選擇性採樣

並列摘要

Lung cancer is one of the most common causes of death today. For cancer, early detection and treatment is the best way to cure. However, it is time-consuming to review hundreds of computed tomography (CT) images for nodule diagnosis in each patient. In recent, the convolution neural network (CNN) had been used in medical imaging for lesion detection, segmentation, and classification and the densely connected convolutional neural network (DenseNet) is one of famous CNN architecture. Hence, in this study, a computer-aided diagnosis (CAD) system, multi-scale densely connected convolutional neural network (MSDN) modified from DenseNet, is proposed to diagnose the nodule as malignant or benign for reducing the physician burden. The major differences between the proposed MSDN and DenseNet are that three different nodule sizes are used as inputs instead of single input and the hyperparameters are reduced by altering the DenseNet structure. Furthermore, in order to improve performance, the pre-training with self-training on two open sources and fine-tuning with selective sampling on our dataset are used during model training and testing. There are 138 benign and 70 malignant cases with pathology proven used in this study. In our experiments, the proposed CAD system achieves an outstanding performance with the accuracy of 93.7%, sensitivity of 93.6%, specificity of 93.6%, and area under the receiver operating characteristic (ROC) curve of 0.973.

並列關鍵字

Lung cancer ； computed tomography ； computer-aided diagnosis ； densely connected network ； self-training ； selective sampling

參考文獻

[1] S. Hamidian, B. Sahiner, N. Petrick, and A. Pezeshk, "3D Convolutional Neural Network for Automatic Detection of Lung Nodules in Chest CT," Proceedings of SPIE--the International Society for Optical Engineering, vol. 10134, p. 1013409, 03/03 2017.

[2] R. L. Siegel, K. D. Miller, and A. Jemal, "Cancer statistics, 2018," CA Cancer J Clin, vol. 68, no. 1, pp. 7-30, Jan 2018.

[3] Q. Song, L. Zhao, X. Luo, and X. Dou, "Using Deep Learning for Classification of Lung Nodules on Computed Tomography Images," in Journal of healthcare engineering, 2017.

[4] H. MacMahon et al., "Guidelines for management of small pulmonary nodules detected on CT scans: a statement from the Fleischner Society," Radiology, vol. 237, no. 2, pp. 395-400, Nov 2005.

[5] N. Tajbakhsh et al., "Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?," IEEE Trans Med Imaging, vol. 35, no. 5, pp. 1299-1312, May 2016.