使用三維膠囊網路於肺部電腦斷層掃描結節偵測

Nodule Detection in Lung CT Using 3-D Capsule Network

林鈺盛(Yu-Sheng Lin)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201902256

摘要

在過去的幾年中，肺癌是最常見也是最致命的癌症之一，早期的發現以及診斷是降低肺癌死亡率以及提升病人存活率的最佳方法。近年來，卷積神經網路作為最常見的深度學習架構，已成為分析醫學影像最先進的方法。例如，密集神經網路以及殘差神經網路利用特徵的重複使用，能夠以少量的參數達到良好的性能。雖然卷積神經網路在眾多計算機視覺領域都展現了強大的性能，它仍存在著一些缺點，例如缺乏空間資訊的情況會嚴重地影響它的辨識能力。一種新穎的架構，膠囊網絡已被證明在建模空間特徵方面是有效的，並且在圖像分類和圖像分割等任務中只需要使用較少的參數。YOLO是一種強大的物件偵測演算法，它通過同時預測邊界框和目標物件的種類來減少計算時間。據我們所知，在近來的文獻中膠囊網路尚未被用作物件偵測演算法中的基底模型。相較於卷積神經網路，膠囊網路在辨識旋轉的物件上具有其優勢，因此可用來提升三維的物件偵測任務的表現。因此，在本研究中提出了一種新的結構，一個三維的膠囊網路並且融合了由殘差神經網路和密集神經網路啟發而得的跳過連接的神經層設計，作為YOLO的基底模型以用於肺結節的偵測。此研究使用了一個開源的數據集LUNA16來驗證系統的性能，結果也證明了相比於現今先進的架構如殘差神經網路以及密集神經網路，經修改過的三維偵測系統可以達到更好的結果。

關鍵字

肺腺癌 ； 電腦斷層掃描 ； 電腦輔助檢測 ； 膠囊網路

並列摘要

In the past several years, lung cancer is the most common and one of the deadliest cancer. Early detection and diagnosis is the best way to reduce the lung cancer mortality rate and improve survival probability. Recently, the convolutional neural network (CNN), the most common architecture of deep learning, have become the state-of-the-art method in medical image analysis. For example, densely connected network (DenseNet) and residual network (ResNet) both achieve good performance with a small number of parameters by feature reuse. Although it has shown powerful performance in various computer vision fields, CNN still has some shortcomings such as the lack of spatial orientation information which will seriously affect the CNN's ability to recognize. A novel model, capsule network (CapsNet) had been proved effective in modeling spatial features with much fewer parameters in image classification and image segmentation. In addition, You Only Look Once (YOLO) is a powerful method for object detection tasks, which reduces computation time by simultaneously predicting bounding boxes and the class of target objects. To our knowledge, the CapsNet is not yet used as a base model for object detection in the current literature. The CapsNet has the ascendancy over CNN in the rotation of objects and would enhance the performance of 3-D object detection tasks. Therefore, in this study, a novel architecture, 3-D CapsNet with YOLO-like detection structure based on the capsule structure integrating skip connection inspired by ResNet and DenseNet, is proposed for lung nodule detection. In this architecture, the not-so-deep linear capsule network (s-Caps) altered based on the replacement of convolutional layer by CapsNet structure is also used to promote the detection performance. For evaluating system performance, the open dataset, Luna16, is used and the results show that the modified 3-D YOLO detection architecture can achieve better performance than other state-of-the-art approaches such as ResNet and DenseNet.

並列關鍵字

Lung cancer ； computed tomography ； computer-aided detection ； capsule network

參考文獻

[1] 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.

[2] R. L. Siegel, K. D. Miller, and A. Jemal, "Cancer statistics, 2019," CA Cancer J Clin, vol. 69, no. 1, pp. 7-34, Jan 2019.

[3] F. Girvin and J. P. Ko, "Pulmonary nodules: detection, assessment, and CAD," AJR Am J Roentgenol, vol. 191, no. 4, pp. 1057-69, Oct 2008.

[4] C. Jacobs, E. M. van Rikxoort, K. Murphy, M. Prokop, C. M. Schaefer-Prokop, and B. van Ginneken, "Computer-aided detection of pulmonary nodules: a comparative study using the public LIDC/IDRI database," Eur Radiol, vol. 26, no. 7, pp. 2139-47, Jul 2016.

[5] Y. Bar, I. Diamant, L. Wolf, S. Lieberman, E. Konen, and H. Greenspan, "Chest pathology detection using deep learning with non-medical training," in 2015 IEEE 12th International Symposium on Biomedical Imaging (ISBI), 2015, pp. 294-297.