電腦斷層掃描肺腫瘤良惡性判別之深度學習影像特徵擷取

Discrimination of Benign and Malignant Pulmonary Tumors in Computed Tomography with Deep Learning Feature Extraction

陳稜鎔(Leng-Jung Chen)

指導教授 ： 陳中明

國立臺灣大學/工學院/醫學工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201804220

摘要

根據衛服部於民國106年統計，癌症高居台灣十大死因之首已達36年，其中又以肺癌為最常見之癌症死因。通過電腦斷層掃描進行肺癌篩檢，有助於提高早期肺癌之診斷率，達到早期治療之目的，提高存活率。肺結節良惡性之評估為肺癌篩檢中重要的一環，肺結節的惡性程度將會影響患者手術決策，於CT影像中發現肺結節時，如何有效的評估肺結節良惡性相當重要。
放射科醫師可依據病患之肺結節惡性程度差異，擬定最佳的臨床治療決策計畫。但惡性程度評估為放射科醫師之主觀判斷，根據放射科醫師經驗不同，在判斷結節良惡性時會有觀察者間之差異，且會因疲勞而增加誤判的可能性，藉由電腦輔助診斷系統(CADx)之引進，可提供放射科醫師另一種客觀意見，降低觀察者間之差異，並為臨床決策提供定量分析，從而提高臨床診斷的性能。
CADx於結節良惡性分類演算法可分成機器學習與深度學習，深度學習是由電腦從大量資料進行訓練，依訓練樣本中學習出有效特徵進行分類，可避免因結節分割結果不同，造成機器學習特徵上的差異，深度學習於結節良惡性分析演算法之準確率可達八成以上。然而，與自然影像之深度學習訓練樣本相比，現有結節資料庫樣本數對於深度學習而言相對缺乏。
由於現有結節樣本數對於深度學習而言過於缺乏，為此，本研究提出兩種演算法 (1) Multiple-Window Convolutional neural network (MWCNN)；(2) Simulated Nodule Augmentation (SNA)。藉由肺結節於CT影像之特性，開發深度學習不同之增量方法，用以進行訓練，克服現今結節樣本數不足之問題。
Multiple-Window Convolutional neural network中，本研究將結節之CT影像，與其Lung Window、Abdomen Window、Bone Window、Chest Window影像作為訓練樣本，其目的在於以臨床上所使用之Window type，將人類對於CT影像之先驗知識，提供深度學習特定HU值之範圍，引導深度學習可藉由此些灰階範圍之影像資訊進行分類。另一方面，藉由CT影像dynamic range之特性，使有限的結節資料量中，延伸不同的灰階資訊，產生gray level的augmentation。本研究以Multiple Channel CNN為架構，且逐一增加Window Channel數，研究結果顯示，當Channel數增加時，CNN之效能亦有所增加，以Dicom影像輸入至3層CNN架構中，其Accuracy：0.823 ±0.015，Sensitivity：0.833 ±0.042，Specificity：0.827 ±0.012，當輸入增加至4種Window Channel時，Accuracy可達0.880 ±0.026，Sensitivity為0.840 ±0.020，Specificity為0.920 ±0.040。
本研究提出Simulated Nodule Augmentation生成擬真結節，克服現有結節樣本數不足之問題。結節依放射科醫師所標記之邊界分割後，將其植入正常肺實質VOI中，可額外增加結節訓練樣本數，藉由固定結節並變更多種環境，使深度學習在訓練過程中，刺激深度學習訓練結節模板，產生對抗環境因素之功用，趨以學習結節本身之特徵。加入SNA產生之擬真結節樣本進行訓練，以3層CNN之結果而言，其Accuracy可達0.883 ±0.006，Sensitivity為0.860 ±0.020，Specificity為0.907 ±0.023，以傳統資料增量作為樣本訓練3層CNN之結果 (Accuracy：0.823 ±0.015，Sensitivity：0.833 ±0.042，Specificity：0.827 ±0.012)相比，在相同架構中，增加SNA之訓練樣本之CNN可獲得較佳結果。
為輔助放射科醫師進行結節良惡性評估，克服現有結節樣本數不足之問題，本研究提出Multiple Window Convolutional neural network與Simulated Nodule Augmentation藉由Multiple Window CNN增加Channel數，可額外提供深度學習多樣的灰階資訊；於Simulated Nodule Augmentation中，結節可另外衍伸出更多樣本數進行訓練。由實驗結果顯示，在有限的結節資料量中，Multiple Window Convolutional neural network與Simulated Nodule Augmentation皆可獲得更佳的良惡性分類結果。

關鍵字

肺結節良惡性分類 ； 深度學習 ； 卷積神經網路 ； 資料增量 ； 深度特徵可視化

並列摘要

According to the 2017 report of Ministry of Health and Welfare, cancer has been ranked the first cause of death in Taiwan for the past 36 years, and lung cancer is one of the leading cause of death. Having lung cancer screening by computed tomography can help improve the early diagnosis rate of lung cancer, achieve early diagnosis and early treatment, and improve the survival rate. The discrimination of benign and malignant pulmonary nodules is an important part of lung cancer screening, whose results will affect the patients’ surgical decision. When lung nodules are found in CT images, how to evaluate the benign and malignant pulmonary nodules effectively is very important.
Radiologists can formulate the best clinical treatment plan based on the suspiciousness of malignancy of the patients' lung nodules. However, the suspiciousness of malignancy is assessed by the radiologist's subjective judgment. The development of computer-aided diagnosis (CADx) system can provide objective opinions for radiologists, reduce the influences of interobservers, and provide quantitative analysis for clinical decisions.
Methods used by CADx system can be classified into machine learning and deep learning in the classification of benign and malignant nodules. Deep learning is performed from a large amount of data, and the effective features are extracted by the training samples. Deep learning can avoid the difference of hand-craft features caused by the difference of nodule segmentation result. Nowadays, the accuracy of deep learning algorithm in the nodule benign and malignant analysis is up to 80%. However, the number of nodule samples is relatively less than the training data of natural images.
Since the number of nodule samples is sparse for deep learning, this study proposed two algorithms (1) Multiple-Window Convolutional neural network (MWCNN); (2) Simulated Nodule Augmentation (SNA). With the characteristics of lung nodules in CT images, we proposed the two different algorithms of deep learning for augmentation and overcome the problem of insufficient nodules samples.
In Multiple-Window Convolutional neural network, this study used CT images, Lung Window, Abdomen Window, Bone Window, and Chest Window images as deep learning. With the prior knowledge of radiologists, we could guide deep learning to train classifier by these specific range of HU values. On the other hand, the limited amount of nodule data was extended by the dynamic range of the CT image. The different grayscale information was extended to generate the gray level augmentation. In this study, the multiple channel CNN was used as the deep learning architecture, and the number of window channels was increased one by one. When the number of channels increases, the performance of CNN also increases. The accuracy of the 3-layer CNN trained by dicom images was 0.823 ± 0.015, sensitivity: 0.833 ± 0.042, specificity: 0.827 ± 0.012. When the input was increased to 4 kinds of window channel, the accuracy was 0.880 ± 0.026, sensitivity was 0.840 ± 0.020, and the specificity was 0.920 ± 0.040.
Simulated Nodule Augmentation algorithm generate simulated nodules to increase training data for deep learning and overcome the problem of insufficient number of nodule samples. After the nodules were segmented according to the boundary annotation by the radiologists, they were implanted into the normal lung parenchymal VOI. These simulated nodules could additionally increase the number of training samples. With the nodules applied various environments, the deep learning model could focus the characteristics of nodule and ignore the influences of background environment during the training process. The accuracy was 0.823 ± 0.015, sensitivity was 0.833 ± 0.042, and specificity: 0.827 ± 0.012 in CNN trained by traditional augmentation data. In the same deep learning architecture, the CNN trained by the data added SNA samples could achieve better performance, the accuracy was 0.883 ± 0.006, the sensitivity was 0.860 ± 0.020, and the specificity was 0.907 ± 0.023
In order to assist the radiologist to evaluate the benign and malignant nodules and overcome the problem of insufficient number of nodule samples, this study proposed Multiple Window Convolutional neural network and Simulated Nodule Augmentation. Multiple Window Convolutional neural network could increase the number of channels, which can provide more grayscale information for deep learning. With Simulated Nodule Augmentation, the training data of nodule could be extended more samples for training. In the limited amount of nodule data, the results show that Multiple Window Convolutional neural network and Simulated Nodule Augmentation could obtain better performance of benign and malignant classification.

並列關鍵字

Discrimination of benign and malignant pulmonary nodule ； deep learning ； convolutional neural network ； data augmentation ； deep learning feature visualization

參考文獻

[1] 105年國人死因統計結果- 衛生福利部。民國106年06月19日。檢自 https://www.mohw.gov.tw/cp-16-33598-1.html

[2] Siegel, Rebecca L., Kimberly D. Miller, and Ahmedin Jemal. "Cancer Statistics, 2017." CA: A Cancer Journal for Clinicians 67.1 (2017): 7-30.

[3] Goldstraw, Peter, et al. "The IASLC lung cancer staging project: proposals for revision of the TNM stage groupings in the forthcoming (eighth) edition of the TNM classification for lung cancer." Journal of Thoracic Oncology 11.1 (2016): 39-51.

[4] Bach, Peter B., et al. "Benefits and harms of CT screening for lung cancer: a systematic review." Jama 307.22 (2012): 2418-2429.

[5] Camarlinghi, Niccolò. "Automatic detection of lung nodules in computed tomography images: training and validation of algorithms using public research databases." The European Physical Journal Plus 128.9 (2013): 110.