臺灣博碩士論文加值系統

骨折（Fracture）是一種常見的創傷性疾病，在國際上具有高度的流行性。特別是手腕部位的骨折，往往是由於跌倒時用手腕支撐而導致的。隨著年齡的增長，骨質疏鬆的問題在高齡者中更為普遍，這使得手腕骨折的風險大大增加。然而，手腕部位的骨骼結構相對複雜，這常常導致放射科醫師在診斷時出現誤判的情況，進而使病情惡化。因此本研究的目標是開發一個完善的電腦輔助診斷系統，協助醫師更準確地診斷手腕骨折，提高診斷效率並降低誤診的風險。
本研究透過Inception及EfficientNet多尺度的特性；DenseNet能夠避免遺失細微特徵的特性制定深度學習模型架構，本研究將整體系統分為兩階段式的深度學習模型架構，第一個階段的網路模型專注於將X光影像根據拍攝角度進行分類，使X光影像分類為左橈骨、右橈骨及側橈骨三種不同的拍攝角度影像，這有助於系統更好地理解和處理不同角度下的影像資訊。第二階段的網路模型用於專注於辨識手腕部位是否存在骨折，首先將影像通過前處理使其能夠專注於手腕的區域並增強亮度不足的區域，且通過小角度資料擴增方式將資料量擴增為原來的兩倍，使其能夠彌補前處理可能造成的誤差角度，然後通過模型堆疊方式將不同網路模型整合辨識手腕骨折並準確偵測骨折可能的區域位置。
研究結果顯示，在第一階段中，使用較淺層且多尺度的Inception-V1深度學習模型，就有較高的分類任務表現，可以得到98.04%的準確率，並且在左橈骨及右橈骨的辨識率都達到100%。在第二階段中，當使用複雜度較高的EfficientNet-B2深度學習模型，才有較高的骨折辨識表現，且在引入了第一階段的拍攝角度分類任務後，其整體模型的性能有著顯著的提升，Cohen’s kappa從72.27%上升至75.11%，可以證明在不同角度辨識骨折的任務具有差異性。然而在其基礎上將不同角度辨識骨折的任務通過模型堆疊融合Inception-V3、EfficientNet-B2及DenseNet-201模型使整體模型的辨識能力更上一層，得到了整體的正確率為88.63%、精確度為95.76%、靈敏度為80.71%、特異度為96.47%、F1 Score為87.59%及Cohen’s kappa為77.26%的數據表現。和其他研究方法比較，結果表示本研究所提出的方法能更精準地辨識手腕骨折及定位手腕骨折位置，提高了診斷的效率和可靠性。

Fractures are a common traumatic disease with a high prevalence internationally. Fractures of the wrist, in particular, are often caused by falling and using the wrist for support. With age, osteoporosis is more prevalent in the elderly, which greatly increases the risk of wrist fractures. However, the complexity of the bone structure in the wrist region leads to misdiagnosis by radiologists, which can lead to deterioration of the condition. Therefore, the goal of this study is to develop a comprehensive computer-aided diagnosis system to help physicians diagnose wrist fractures more accurately, improve diagnostic efficiency and reduce the risk of misdiagnosis.
In this study, the deep learning model architecture is formulated through the multi-scale characteristics of Inception and EfficientNet, and the characteristics of DenseNet that can avoid the loss of subtle features, and the whole system is divided into a two-stage deep learning model architecture in this study. The first stage of the network model focuses on the classification of the X-ray images according to the shooting angle, so that the X-ray images are classified into three different shooting angle images, namely, left radius, right radius and lateral radius. The first phase of the network model focuses on categorizing the X-ray images according to the camera angles so that the X-ray images are categorized into left radius, right radius, and lateral radius, which helps the system to better understand and process the image information in different angles. In the second stage, the network model is used to focus on identifying whether there is a fracture in the wrist area. Firstly, the image is pre-processed so that it can focus on the wrist area and enhance the areas that are not bright enough, and the amount of data is doubled through the small-angle data augmentation method to make up for the error angle that may be caused by the pre-processing, and then different network models are integrated to identify the fracture of the wrist through the model stacking method. Then, different network models are integrated through model stacking to recognize the wrist fracture and accurately detect the possible location of the fracture area.
The results of the study show that in the first stage, when using the shallower and multi-scale Inception-V1 deep learning model, the performance of the classification task is higher, and an accuracy of 98.04% can be obtained, and the recognition rate of left radius and right radius reaches 100%. In the second stage, when the EfficientNet-B2 deep learning model with higher complexity is used, it has a higher fracture recognition performance, and after the introduction of the camera angle classification task in the first stage, its overall model performance has a significant improvement, and Cohen's kappa increases from 72.27% to 75.11%, which can be Cohen's kappa increased from 72.27% to 75.11%, which can prove that the task of recognizing fracture at different angles is different. However, the task of recognizing fractures at different angles is further enhanced by fusing the Inception-V3, EfficientNet-B2 and DenseNet-201 models through model stacking, and the overall accuracy of 88.63%, precision of 95.76%, sensitivity of 80.71%, specificity of 96.47%, F1 Score of 87.59% and Cohen's kappa of 77.26%. When comparing with the state-of-the-art methods in the literature, the method proposed in this study can recognize wrist fracture and locate the wrist fracture position more accurately, which demonstrates the efficiency and reliability of diagnosis.

目錄
誌謝 i
摘要 ii
Abstract iv
目錄 vi
圖目錄 ix
表目錄 xi
第一章 緒論 1
1.1研究動機 1
1.2研究目的 2
1.3研究架構 3
第二章 研究背景 4
2.1手腕骨頭與骨折 4
2.2手腕X光拍攝方式 5
2.3深度學習 6
2.4相關文獻探討 7
第三章 研究方法 14
3.1資料庫介紹 14
3.2影像前處理 17
3.2.1外框對齊分割 18
3.2.1.1影像二值化 19
3.2.1.2形態學運算 20
3.2.1.3主成份分析(PCA) 21
3.2.2去標註去背景 22
3.2.2.1直方圖均衡化 (Histogram Equalization) 24
3.2.3去外框並骨頭對齊 25
3.2.4對比度增強 26
3.3資料擴增 28
3.4分類模型架構 29
3.4.1 Inception模塊 30
3.4.2 密集網路模型(DenseNet model) 34
3.4.3 EfficientNet model 37
3.4.4 堆疊法 (Stacking) 38
3.4.4.1 邏輯迴歸 (Logistic Regression) 39
3.4.4.2 線性支持向量機 (Linear SVM) 40
3.4.4.3高斯徑向基函數支持向量機 (RBF SVM) 40
3.4.4.4 K近鄰 (KNN) 40
3.4.4.5 高斯樸素貝葉斯 (GaussianNB) 41
3.4.4.6 決策樹 (Decision Tree) 41
3.4.4.7 隨機森林 (Random Forest) 41
3.4.4.8 AdaBoost 42
3.4.4.9 Gradient Boosting 42
3.4.4.10 XGBoost 42
3.5損失函數 43
3.6模型可視化 43
3.7硬體設備與軟體環境 44
3.7.1硬體設備 44
3.7.2軟體環境 44
第四章 研究結果與討論 45
4.1前處理之實驗結果 45
4.2評估指標 46
4.2.1二分類之混淆矩陣 46
4.2.2三分類之混淆矩陣 48
4.3訓練參數設定 50
4.4模型實驗結果 51
4.4.1橈骨定位實驗結果 51
4.4.2手腕骨折實驗結果 53
4.4.2.1同一模塊及模型之不同層數架構對骨折分類的影響 53
4.4.2.2影像前處理與資料擴增對骨折分類的影響 54
4.4.2.3拍攝角度對骨折分類的影響 56
4.4.2.4不同變因的實驗結果比較 58
4.4.2.5模型堆疊對骨折分類的影響 59
4.4.2.6模型可視化結果 61
4.5相關文獻比較 65
第五章 結論與未來展望 67
5.1結論 67
5.2未來展望 68
參考文獻 69

[1]臺灣一年超過10萬人因骨折就醫 Available:https://www.advisers.com.tw/?p=11832(2024, 06, 28)
[2]台灣骨折率亞洲之冠　長者1年內死亡率36%「比癌症更致命」
Available:https://www.healthnews.com.tw/article/59624(2024, 06, 28)
[3]三總醫世代電子報第244期
Available:https://enews.tsgh.ndmctsgh.edu.tw/edm/content_detail.aspx?eid=548(2024, 06, 28)
[4]X光片下的灰白，如何利用X光片檢查關節傷害
Available:https://www.sem.org.tw/EJournal/Detail/219(2024, 06, 28)
[5]Yang, S., Yin, B., Cao, W., Feng, C., Fan, G., & He, S. (2020). Diagnostic accuracy of deep learning in orthopaedic fractures: a systematic review and meta-analysis. Clinical radiology, 75(9), 713-e17.
[6]Hardalaç, F., Uysal, F., Peker, O., Çiçeklidağ, M., Tolunay, T., Tokgöz, N., ... & Mert, F. (2022). Fracture detection in wrist X-ray images using deep learning-based object detection models. Sensors, 22(3), 1285.
[7]Bhat, A. K., Kumar, B., & Acharya, A. (2011). Radiographic imaging of the wrist. Indian journal of plastic surgery: official publication of the Association of Plastic Surgeons of India, 44(2), 186.
[8]Trauma X-ray - Upper limb
Available:https://www.radiologymasterclass.co.uk/tutorials/musculoskeletal/x-ray_trauma_upper_limb/wrist_trauma_x-ray(2024, 06, 28)
[9]深度學習是什麼？
Available:https://ithelp.ithome.com.tw/articles/10264238?sc=iThelpR(2024, 06, 28)
[10]維基百科深度學習
Available:https://zh.wikipedia.org/zh-tw/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0(2024, 06, 28)
[11]CNN：卷積深度神經網路的解釋方法
Available:https://ithelp.ithome.com.tw/articles/10332039?sc=rss.iron(2024, 06, 28)
[12]激活函數
Available:https://ithelp.ithome.com.tw/m/articles/10276865(2024, 06, 28)
[13]Batch Normalization 介紹
Available:https://medium.com/ching-i/batch-normalization-%E4%BB%8B%E7%B4%B9-135a24928f12(2024, 06, 28)
[14]Rashid, T., Zia, M. S., Meraj, T., Rauf, H. T., & Kadry, S. (2023). A minority class balanced approach using the DCNN-LSTM method to detect human wrist fracture. Life, 13(1), 133.
[15]Alammar, Z., Alzubaidi, L., Zhang, J., Li, Y., Lafta, W., & Gu, Y. (2023). Deep transfer learning with enhanced feature fusion for detection of abnormalities in x-ray images. Cancers, 15(15), 4007.
[16]Rajpurkar, P., Irvin, J., Bagul, A., Ding, D., Duan, T., Mehta, H., ... & Ng, A. Y. (2017). Mura: Large dataset for abnormality detection in musculoskeletal radiographs. arXiv preprint arXiv:1712.06957.
[17]Chollet, F. (2017). Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1251-1258).
[18]Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2017, February). Inception-v4, inception-resnet and the impact of residual connections on learning.In Proceedings of the AAAI conference on artificial intelligence (Vol. 31, No. 1).
[19]Joshi, D., & Singh, T. P. (2022, December). Novel Use of Deep Convolution Architecture Pre-Trained on Surface Crack Dataset to Localize and Segment Wrist Bone Fractures. In 2022 11th International Conference on System Modeling & Advancement in Research Trends (SMART) (pp. 1308-1313). IEEE.
[20]Yang, F., & Ding, B. (2020, September). Computer aided fracture diagnosis based on integrated learning. In 2020 IEEE 3rd International Conference on Information Systems and Computer Aided Education (ICISCAE) (pp. 523-527). IEEE.
[21]Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90.
[22]Pearson, K. (1901). Principal components analysis. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 6(2), 559.
[23]Ananda, A., Ngan, K. H., Karabağ, C., Ter-Sarkisov, A., Alonso, E., & Reyes-Aldasoro, C. C. (2021). Classification and visualisation of normal and abnormal radiographs; a comparison between eleven convolutional neural network architectures. Sensors, 21(16), 5381.
[24]Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1-9).
[25]Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
[26]Iandola, F. N., Han, S., Moskewicz, M. W., Ashraf, K., Dally, W. J., & Keutzer, K. (2016). SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size. arXiv preprint arXiv:1602.07360.
[27]He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
[28]Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2818-2826).
[29]Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708).
[30]Pizer, S. M. (1990, May). Contrast-limited adaptive histogram equalization: Speed and effectiveness stephen m. pizer, r. eugene johnston, james p. ericksen, bonnie c. yankaskas, keith e. muller medical image display research group. In Proceedings of the first conference on visualization in biomedical computing, Atlanta, Georgia (Vol. 337, p. 1).
[31]Oh, J., Hwang, S., & Lee, J. (2023). Enhancing X-ray-Based Wrist Fracture Diagnosis Using HyperColumn-Convolutional Block Attention Module. Diagnostics, 13(18), 2927.
[32]Radosavovic, I., Kosaraju, R. P., Girshick, R., He, K., & Dollár, P. (2020). Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10428-10436).
[33]Zoph, B., Vasudevan, V., Shlens, J., & Le, Q. V. (2018). Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 8697-8710).
[34]Tan, M., & Le, Q. (2019, May). Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning (pp. 6105-6114). PMLR.
[35]Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708).
[36]Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... & Houlsby, N. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.
[37]Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., ... & Guo, B. (2021). Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 10012-10022).
[38]Woo, S., Park, J., Lee, J. Y., & Kweon, I. S. (2018). Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV) (pp. 3-19).
[39]Hržić, F., Štajduhar, I., Tschauner, S., Sorantin, E., & Lerga, J. (2019). Local-entropy based approach for X-ray image segmentation and fracture detection. Entropy, 21(4), 338.
[40][圖像處理] 二值化閾值自動化篩選 - Otsu, 多重門檻值, 直方圖
Available:https://medium.com/@mingjiehsu/%E4%BA%8C%E5%80%BC%E5%8C%96%E9%96%A5%E5%80%BC%E8%87%AA%E5%8B%95%E5%8C%96%E7%AF%A9%E9%81%B8-otsu-%E5%A4%9A%E9%87%8D%E9%96%80%E6%AA%BB%E5%80%BC-%E7%9B%B4%E6%96%B9%E5%9C%96-345aff032e0f(2024, 06, 28)
[41]形態學操作(膨脹,腐蝕,開操作,閉操作)
Available:https://www.twblogs.net/a/5b846e6d2b71775d1cd0b728(2024, 06, 28)
[42]Ostu, N. (1979). A threshold selection method from gray-level histograms. IEEE Trans SMC, 9, 62.
[43][OpenCV] 淺談直方圖均衡化Histogram Equalization、AHE均衡、CLAHE均衡(2024, 06, 28)
Available:https://medium.com/@cindylin_1410/%E6%B7%BA%E8%AB%87-opencv-%E7%9B%B4%E6%96%B9%E5%9C%96%E5%9D%87%E8%A1%A1%E5%8C%96-ahe%E5%9D%87%E8%A1%A1-clahe%E5%9D%87%E8%A1%A1-ebc9c14a8f96
[44]何謂資料增強
Available:https://aws.amazon.com/tw/what-is/data-augmentation/(2024, 06, 28)
[45]Wolpert, D. H. (1992). Stacked generalization. Neural networks, 5(2), 241-259.
[46]卷積神經網絡 CNN 經典模型 — GoogleLeNet、ResNet、DenseNet with Pytorch code
Available:https://medium.com/ching-i/%E5%8D%B7%E7%A9%8D%E7%A5%9E%E7%B6%93%E7%B6%B2%E7%B5%A1-cnn-%E7%B6%93%E5%85%B8%E6%A8%A1%E5%9E%8B-googlelenet-resnet-densenet-with-pytorch-code-1688015808d9(2024, 06, 28)
[47]Bishop, C. M., & Nasrabadi, N. M. (2006). Pattern recognition and machine learning (Vol. 4, No. 4, p. 738). New York: springer.
[48]James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning (Vol. 112, p. 18). New York: springer.
[49]Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626).
[50]Chawla, N., & Bedwa, M. (2022, October). Optimized Ensemble Learning Technique on Wrist Radiographs using Deep Learning. In 2022 2nd International Conference on Technological Advancements in Computational Sciences (ICTACS) (pp. 152-159). IEEE.