預訓練與轉移卷積神經網路於分類胸部肺炎X光之效能評估

Investigation of Classified Performances between Pre-Trained and Transferred CNNs for Pneumonia Chest X-ray Images

王豪善(Hao-Shan Wang)

指導教授 ： 謝文權

共同指導教授 ： 陳泰賓(Tai-Been Chen)

義守大學/醫學科技學院/生物科技學系/碩士(2022年)

摘要

研究目的: 卷積神經網路(Convolutional Neural Networks, CNNs)被廣泛使用於影像分類，使用方式包含:使用預訓練模型(Pre-Trained)進行影像特徵萃取後，經由機器學習分類器分類及透過微調預訓練模型參數後，以轉移學習(Transferred Learning)方式重新訓練後分類，基於預訓練及轉移學習兩種方法皆廣泛被使用，故比較兩者間之效能。
材料與方法: 本研究採用回顧性平衡分組實驗設計，收集網路公開COVID-19(Coronavirus Disease)、非COVID-19肺炎及正常之胸部X光影像之資料庫各300張。實驗因子為影像套色處理包括灰階影像及二種偽色影像。採隨機選取70%影像用於訓練模型；30%用於驗證深度學習模型。預訓練與轉移學習CNN模型包含Alexnet、VGG19、ResNet18、ResNet50、Densenet201及Darknet53；其中預訓練CNN主要用於影像特徵萃取再經由SVM(Support Vector Machine)、KNN(K-Nearest Neighbors)及LR(Logistic Regression)建立分類模型；而轉移學習CNN模型主要是經過修改全連接層(Fully Connection Layer)之分組個數並經過訓練後，再經由驗證而得到之模型。
結果: 轉移學習模型分類結果以Darknet53使用灰階影像具最高準確度，Accuracy、Sensitivity、Precision、F-score、Kappa值為0.96、0.96、0.96、0.96、0.93；預訓練模型結合分類器以Darknet53結合SVM使用灰階影像具最高準確度，Accuracy、Sensitivity、Precision、F-score、Kappa值為0.96、0.96、0.96、0.96、0.94，結果顯示預訓練模型結合分類器在較少的訓練時間下，達到略高於轉移學習之效能。
結論: 兩組偽色影像分類效能略低於灰階影像，又以Jet影像效能優於Parula影像。此外，預訓練模型結合機器學習分類器方法優於以轉移學習方式使用CNN模型，又以Darknet53適合用於特徵萃取，SVM作為萃取特徵之分類器。

關鍵字

新冠肺炎 ； 胸部X光 ； 深度學習 ； 分類器

並列摘要

Purpose: Convolutional Neural Networks (CNNs) are widely used in image classification. CNN was used to extract image features and classify categories by the machine learning approach. Fine-tuning parameters for the pre-trained CNN, which was named as transferred learning schema, were used to classify categories. Based on two methods of pre-trained and transfer learning are widely used, comparing the classification performance of the two methods.
Methods and Materials: In this study, a retrospective and balanced grouping experimental design to collect Coronavirus Disease (COVID-19), non-COVID-19 pneumonia, and normal CXR 300 images in each category from the open database on the Internet. The experimental factors are color processing, including original grayscale, and two pseudo-color processing. Random split 70% for training; the rest of 30% to validate the deep learning model. Pre-trained transfer learning models include Alexnet, VGG19, ResNet18, ResNet50, Densenet201, and Darknet53. The pre-trained model is a major used for feature extraction, then establishing classification model by SVM (Support Vector Machine), KNN (K-Nearest Neighbors), and LR (Logistic Regression). The transfer learning models are mainly obtained by adjusting the number of groups of the fully connected layer, retraining, and validating.
Results: After transfer learning, the Darknet53 model generated the highest accuracy with grayscale images. Accuracy, sensitivity, precision, F-score, and Kappa value were 0.96, 0.96, 0.96, 0.96, and 0.93. The Darknet53 combined with SVM generated the highest accuracy with grayscale images. Accuracy, sensitivity, precision, F-score, and Kappa value were 0.96, 0.96, 0.96, 0.96, and 0.94. The classification result shows that pre-trained models combined with a machine learning classifier performed better than transfer learning approaches.
Conclusion: The performance of the classification with grayscale images was better than using pseudo-color images (i.e., Jet and Parula Colormaps). In addition, the performance of classification by the pre-trained model combined with the machine learning classifier was better than the transferred learning method. Hence, the Darknet53 was suitable for feature extraction, and SVM was suitable to be a classifier in this study.

並列關鍵字

COVID-19 ； Chest X-ray ； Deep Learning ； Classifier

參考文獻

[1] M. H. Chowdhury, T. Rahman, A. Khandakar, R. Mazhar, M. A. Kadir, Z. B. Mahbub, K. R. Islam, M. S. Khan, A. Iqbal, N. A. Emadi, M. B. I. Reaz, and M. T. Islam, "Can AI help in screening viral and COVID-19 pneumonia?." IEEE Access 8, 132665-132676, 2020.

[2] L. Wang, Z. Q. Lin, and A. Wang, "Covid-net: A tailored deep convolutional neural network design for detection of covid-19 cases from chest x-ray images." Scientific Reports, 1-12, 2020.

[3] Y. Oh, S. Park, and J. C. Ye, "Deep learning covid-19 features on cxr using limited training data sets." IEEE transactions on medical imaging 39.8, 2688-2700, 2020.

[4] M. Ozsoz, A. U. Ibrahim, S. Serte, F. A. Turjman, and P. S. Yakoi, "Viral and bacterial pneumonia detection using artificial intelligence in the era of COVID-19." , 2020.

[5] M. P. Karthikeyan, "An Efficient Deep Learning Approach to Pneumonia Classification in Healthcare." International Journal of Advanced Research in Science & Technology, 2020.