使用深度學習方法對肺癌患者之鎝-99m-亞甲基雙磷酸鹽骨掃描影像進行骨轉移分類

Classification for Bone Metastasis of 99mTc-MDP Bone Scan Images in Patient with Lung Cancer Via Deep Learning Methods

吳宜臻(Yi-Chen Wu)

指導教授 ： 陳泰賓

共同指導教授 ： 杜維昌(Wei-Chang Du)

義守大學/智慧科技學院/資訊工程學系/博士(2022年)

摘要

動機與目的： 99mTc-MDP(methylene diphosphonate)骨掃描是目前診斷癌症骨轉移的主要影像工具。然而，受限於影像的解析度，檢查的靈敏度及特異性，以及目視判讀和觀察者間變異性的影響，假陽性及假陰性仍然是診斷骨轉移的重要問題。因此，本研究將應用深度學習演算法來構建用於骨轉移的人工智能分類模型。
材料與方法：本回顧性研究將收集自2004年8月至2013年12月間，義大醫院肺癌患者之全身骨掃描影像共1474筆，並分為骨轉移影像和無骨轉移影像兩類，各含487筆及987筆。影像分為正面、正面加背面及兩種灰階正面加背面三組數據集，透過七種卷積神經網絡 (Convolution Neural Network, CNN) 模型進行影像特徵萃取，再由三種機器學習包括支持向量機(Support Vector Machine, SVM)、單純貝氏分類器(Nave Bayes)及邏輯斯特回歸(Logistic Regression)建立分類模型，其中90%用於訓練，10%用於測試模型，以10折交叉驗證(10-fold cross validation) 測試模型的性能。最後以測試組的準確性、靈敏度、特異性及 Kappa 一致性統計量評估模型效能。
結果：在三組影像數據集中，由AlexNet與SVM組合的分類模型，都顯示出最好的骨轉移分類效能。其中包含兩種灰階正面加背面影像的數據集， 以AlexNet與SVM組合的分類模型顯示出最佳的分類效率，其準確度為 0.814，Kappa 值為 0.555。
結論：在本研究中，使用預訓練CNN進行特徵提取與機器學習作為分類器，可以成功的對肺癌骨轉移的全身骨掃描影像進行分類。包含兩種灰階正面加背面影像的數據集，可以在肺癌骨轉移的骨掃描影像中提供較多特徵而有較好的分類效能。

關鍵字

深度學習 ； 影像分類 ； 骨掃描 ； 骨轉移 ； 肺癌

並列摘要

Motivation and purpose: 99mTc-MDP (methylene diphosphonate) bone scan is the most important imaging tool for the diagnosis of bone metastasis. However, because of the resolution of the image, sensitivity and specificity of the study, bias from visual interpretation and inter-observer variability, false positives and false negatives are still important issues in the diagnosis of bone metastasis. Therefore, this study will apply deep learning algorithms to build an artificial intelligence classification model for bone metastasis.
Materials and methods: This retrospective study collect 1,474 whole body bone scan images of lung cancer patients at E-Da Hospital from August 2004 to December 2013. They are labeled as bone metastasis images and non-bone metastasis images, each containing 487 and 987 images. The image is divided into three data sets: (1) anterior view image (AP), (2) anterior and posterior view image (AP+PA), and (3) two grey scale anterior and posterior view image (Dual AP+PA). Feature extraction is performed by seven Convolution Neural Network (CNN) models with three types of machine learning including support vector machines (Support Vector Machine, SVM), Nave Bayes (Nave Bayes) and Logistic Regression (Logistic Regression) to establish classification models. Of these, 90% is used for training and 10% is used for testing the model. Test the performance of the model with 10-fold cross validation. Finally, the accuracy, sensitivity, specificity and Kappa value of the test group were used to evaluate the performance of the model.
Results: The AlexNet + SVM classifier has better performance than those of the presented methods in all data sets. The data set 3 (Dual AP+PA) with two grey scale anterior and posterior view image shows the best classification efficiency by pre-trained CNN and machine learning approach. The accuracy is 0.814 and the Kappa value is 0.555.
Conclusions: In this study, pre-trained CNN for feature extraction with machine learning as a classifier is a suitable method for classification of bone metastasis in lung cancer in whole body bone scans images. Data set with two grey scale anterior and posterior view image has the most characteristics of bone metastasis in lung cancer in bone scan image for classification.

並列關鍵字

deep learning ； image classification ； bone scan ； bone metastasis ； lung cancer

參考文獻

[1] The Cancer Registration Report of the Ministry of Health and Welfare in 2018, 2020.

[2] American Cancer Society. Key statistics for lung cancer. Published 2017. Accessed October 23, 2017.

[3] Siegel RL, Miller KD, Jemal A. Cancer statistics, 2015. CA Cancer J Clin. 2015;65(1):5-29. doi:10.

[4] Alberg AJ, Brock MV, Ford JG, Samet JM, Spivack SD. Epidemiology of lung cancer: diagnosis and management of lung cancer, 3rd ed: American College of Chest Physicians evidence-based clinical practice guidelines. Chest. 2013;143(5)(suppl):e1Se29S.

[5] Jonas D, Reuland DS, Reddy SM, et al. Screening for Lung Cancer With Low-Dose Computed Tomography: An Evidence Review for the US Preventive Services Task Force. Evidence Synthesis No. 198. Agency for Healthcare Research and Quality; 2021.