臺灣博碩士論文加值系統

臨床超音波影像品質通常為主觀性評估，目前為止超音波影像品質評估標準無明確定義；因此，本研究使用預訓練模型ResNet50以及傳統統計分析方法針對腹部超音波影像品質進行分析。
超音波儀器為SonoSite 180+及扇形探頭頻率為2-5MHz，掃描人體腹部假體(E-130328-368T)進行實驗，掃瞄範圍為肋緣以下前腹部主要於影像中可觀察到肝、膽、胰、脾、腎之位置為主，造影深度共五種深至淺分別為腹部皮下22cm、20cm、17cm、15cm、12cm，共268張影像，其中包含134張高品質影像及134張低品質影像；將影像分為訓練集140張、驗證集70張及測試集70張；計算每張影像訊雜比(Signal-to-noise ratio, SNR)、對比值(Contrast ratio, CR)及總變異量(Total Variation, TV)量化影像品質，進行邏輯斯特迴歸(Logistic Regression)建立分類模型使用；採監督式學習(Supervised learning)，參考卷積神經網路之ResNet50架構以轉移學習(Transfer Learning)方式進行訓練，使用循序搜尋法(Linear Search)調整參數，將驗證集準確度最高之參數輸出作為後續ResNet50建模之用；一為直接使用ResNet50分類影像品質；二為利用ResNet50輸出之影像特徵作為預測變量，建立支持向量機(Support vector machine, SVM)影像品質分類器。
經由轉移學習後訓練完成之ResNet50直接進行分類測試集影像，分類準確度為98%、靈敏度為96%、特異度為100%；將相同訓練完成之ResNet50模型萃取出的特徵參數輸入至SVM進行分類相同測試集影像，分類準確度為98%、靈敏度為100%、特異度為96%；將每張影像之SNR及TV值量化後，進行邏輯斯特迴歸建立一分類模型，其分類準確度為97%、靈敏度為99%、特異度為94%。
實驗中將ResNet50與ResNet50+SVM兩種分類方式準確度較高；目前超音波影像品質評估方法有限且無明確定義，本研究提出了三種分類超音波影像品質評估方法。

In clinics, the quality of ultrasound image is usually subjective determination. So far, the quality of ultrasound image is not clearly defined proper criteria. In this study, ResNet50 approach and logistic regression model were used to classify the quality of ultrasound images.
The SonoSite 180+ and curve probe with frequency 2-5 MHz were used to be imaging instruments. Experimental ultrasound images were obtained from a human abdominal prosthesis (E-130328-368T). The locations of imaging were along with epigastric region. The organs were including liver, gallbladder, pancreas, spleen and kidney. The depth of imaging were including 22cm, 20cm, 17cm, 15cm and 12cm. A total of 268 images were obtained in which both high-quality and low-quality images were 134 individually. The number of images among training, validation, and testing sets were 140, 70, and 70. Each image was calculated signal-to-noise ratio (SNR), contrast ratio (CR) and total variation (TV). One Logistic Regression (LR) was applied to establish classified model for groups between high-quality and low-quality. Meanwhile, the classifiers between high-quality and low-quality groups were adopted sequence searching schema in order to obtained suitable parameters for ResNet50 via transfer learning procedure on validation set. Next, the ResNet50, a convolutional neural network algorithm, was also utilized to build a classifier for groups. The last approach was combined ResNet50 and support vector machine (SVM) to construct a classifier for groups (ResNet50+SVM). The performance of these classifiers was investigation of accuracy, specificity, and sensitivity on testing set.
The accuracy, sensitivity, and specificity were 97%, 99%, and 94% provided by LR model with predictors SNR and TV on whole sample. The accuracy, sensitivity, and specificity were 98%, 96%, and 100% provided by ResNet50 model on testing sample. The accuracy, sensitivity, and specificity were 98%, 100%, and 96% provided by ResNet50+SVM model on testing sample.
In this work, the presented methods were generated recommendable accuracy. Moreover, the performances of ResNet50 and ResNet50+SVM were simpler and more accurate than those of LR. Meanwhile, these presented methods were approved to provide objective judgment for quality of ultrasonic image.

摘 要 I
Abstract II
誌謝 III
目錄 IV
表目錄 V
圖目錄 VI
第一章 緒論 1
1.1前言 1
1.2研究動機與目的 2
第二章 文獻探討 3
第三章 研究材料與方法 7
3.1研究資料來源 7
3.2造影儀器及條件 7
3.3研究流程 8
3.4定義影像品質及分類 9
3.5影像前處理 13
3.6 ResNet50及支持向量機 13
3.7 ResNet50架構及參數 18
3.8 統計分析方法 25
3.9分類成效評估方法 26
第四章 結果 28
4.1 ResNet50模型訓練結果 28
4.2 ResNet50、ResNet50+SVM及邏輯斯特迴歸模型分類結果 29
第五章 結論與未來研究方向 31
5.1結論 31
5.2未來研究方向 31
第六章 研究限制 32
參考文獻 33

[1]Christian MH. Ultrasound Image Quality Assessment: A framework for evaluation of clinical image quality. Medical Imaging. 2010.7629
[2]Kishor K. Objective Image Quality Metrics for Ultrasound Imaging. 2012.
[3]Fukushima K. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biological cybernetics. 1980. 36(4):193-202.
[4]LeCun Y, Bottou L, Bengio Y, and Haffner P. Gradient-based learning applied to document recognition. IEEE. 1998. 86(11):2278-2324.
[5]LeCun Y, Bengio Y, and Hinton G. Deep learning. Nature. 2015. 521(7553):436–444.
[6]Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, and Rabinovich A. Going deeper with convolutions. IEEE conference on computer vision and pattern recognition. 2015:1-9.
[7]Simonyan K, and Zisserman A. Very deep convolutional networks for large-scale image recognition. 2014. arXiv:1409.1556.
[8]Zeiler MD, and Fergus R. Visualizing and understanding convolutional networks. In European conference on computer vision. 2014: 818–833.
[9]Eigen D, Rolfe J, Fergus R, and LeCun Y. Understanding deep architectures using a recursive convolutional network. 2013. arXiv:1312.1847.
[10]Krizhevsky A, Sutskever I, and Hinton GE. Imagenet classification with deep convolutional neural networks. Neural informationprocessing systems. 2012: 1097–1105.
[11]Wen TH, Gasic M, Mrksic N, Su PH, Vandyke D, and Young S. Semantically conditioned lstm-based natural language generation for spoken dialogue systems. 2015. arXiv: 1508.01745.
[12]Chorowski JK, Bahdanau D, Serdyuk D, Cho K, and Bengio Y. Attention-based models for speech recognition. Neural Information Processing Systems. 2015: 577–585.
[13]Yang Z, Yang D, Dyer C, He X, Smola AJ, and Hovy EH, Hierarchical Attention Networks for Document Classification. North American Chapter of the Association for Computational Linguistics: Human Language Technologies.2016:1480-1489.
[14]Agarap AF. A Neural Network Architecture Combining Gated Recurrent Unit (GRU) and Support Vector Machine (SVM) for Intrusion Detection in Network Traffic Data. 10th International Conference on Machine Learning and Computing. 2018: 26-30.
[15]Alalshekmubarak A, and Smith LS. A novel approach combining recurrent neural network and support vector machines for time series classification. 9th International Conference on Innovations in Information Technology. 2013: 42–47.
[16]Hassan TM, Elmogy M, Sallam S. Diagnosis of focal liver diseases based on deep learning technique for ultrasound images. Arabian Journal for Science and Engineering August. 2017. 42(8): 3127–3140.
[17]Wu L, Cheng JZ, Li S, Lei B, Wang T, and Ni D. FUIQA: Fetal ultrasound image quality assessment with deep convolutional networks. IEEE transactions on cybernetics. 2017. 47(5):1336-1347.
[18]Tajbakhsh N, Shin JY, Gurudu SR, Hurst RT, Kendall CB, Gotway MB, and Liang J. Convolutional neural networks for medical image analysis: full training or fine tuning? IEEE transactions on medical imaging. 2016. 35(5):1299-1312.
[19]Agarap A F. An architecture combining convolutional neural network (CNN) and support vector machine(SVM) for image classification. 2019. arXiv:1712.03541v2.
[20]Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift. 2015. arXiv:1502.03167v3.
[21]Tang Y. Deep learning using linear support vector machines.2015. arXiv:1306.0239v4.
[22]He K, Zhang X, Shaoqing, and Sun RJ. Deep residual learning for image recognition. IEEE conference on computer vision and pattern recognition 2016:770-778.
[23]Helou ES, Censor Y, Chen TB, Chern IL, De Pierro AR, Jiang M, Lu HHS. String-averaging expectation-maximization for maximum likelihood estimation in emission tomography. Inverse Problems. 2014. 30(5).
[24]Li X, Zhang S, Zhang Q, Wei X, Pan Y, Zhao J, and Yang F. Diagnosis of thyroid cancer using deep convolutional neural network models applied to sonographic images: a retrospective, multicohort, diagnostic study. The Lancet Oncology. 2019. 20(2):193-201.