臺灣博碩士論文加值系統

本論文提出一種基於 Fast Region-based Convolutional Neural Network (Fast R-CNN) 的方法，用於實現乳癌磁共振成像 (Magnetic Resonance Imaging, MRI) 的自動檢測。本論文選擇The Cancer Image Archive (TCIA) 的 Duke Breast Cancer MRI Dataset 做為資料集，其中包含3120筆乳癌 MRI 影像訓練集，1310 筆乳癌 MRI 影像驗證集。
在檢測流程中的前處理階段，透過影像對比增強的方法強化癌症的特徵，並消除其他不必要的影像資訊，包含器官與乳房的輪廓，讓檢測結果的準確度 (Accuracy) 大幅提升；在建模階段，本論文包含兩次預測階段，分別為第一次預測及第二次預測。在第一次預測階段，利用基於神經網絡 (Neural Network) 模型的方法搜索所有的錨框 (Anchor Boxes)，並選出分數較高的前幾個錨框，其作用類似於快速篩選的概念。不僅能提升檢測速度，更能優先選出可能更具意義的錨框。被選出的錨框再分類成候選預測框，然後根據輪廓特徵對預測框進行修正；在第二次預測階段，本論文將修正後的候選預測框再次使用卷積神經網絡 (Convolutional Neural Network) 模型進行腫瘤辨識，從而再降低偽陽性 (False Positive) 並最終達到更精確的預測效果。
本論文所使用之檢測方法其效率為 0.27 秒可以檢測一張乳癌 MRI 影像。而 MRI 照出單一張影像的時間約為數秒，因此該方法可應用於即時檢測。另外由於 Duke Breast Cancer MRI Dataset 的腫瘤標註格式為每個病例的乳癌 MRI 影像的全身掃描合集中，共用一個註釋框。這導致在實際情況中，對某一張乳癌 MRI 影像中的腫瘤資訊可能很難完全匹配。因此本論文將研究成果分成三個部分呈現，第一部分為Intersection over Union (IoU) 大於 0、精確度 (Precision) 可達到 94.5%、召回率 (Recall) 可達到 89.5%、F-score 可達到 92%；第二部分為 IoU 大於等於0.5，並排除註釋框可能跟腫瘤資訊差異較大的乳癌 MRI 影像，其中包含 197 筆驗證集乳癌 MRI 影像，精確度可達到 83.7%、召回率可達到 87.1%、平均精度 (AP50) 可達到 77.7%、F-score 可達到 85.4%；第三部分為選用IoU 大於等於0.5 的資料，並用 YOLOv7 (You Only Look Once version 7) 預測，其中訓練集有 2365 筆，驗證集有 970 筆資料，精確度為 61.12%、召回率為 75%、AP50 為 72.8%、F-score 為 67.35%。
從研究成果得知，本論文提出的檢測方法優於YOLOv7且能夠檢測出大多數癌症的位置，雖然 AP50 的分數仍有待提高，但本論文提出的方法將成為乳癌 MRI 自動檢測的一個新起點。

This thesis proposes a method based on a Fast Region-based Convolutional Neural Network (Fast R-CNN) for the automated detection of breast cancer in Magnetic Resonance Imaging (MRI). The Cancer Image Archive (TCIA)'s Duke Breast Cancer MRI Dataset is chosen as the dataset, comprising 3120 training samples and 1310 validation samples of breast cancer MRI images.
This thesis uses image contrast enhancement methods to enhance cancer features and remove unnecessary image information, including organ and breast contours, in the preprocessing stage of the detection process. It can significantly increase the accuracy of the detection results. In the modeling stage, this thesis includes two prediction stages, the first prediction and the second prediction. In the first prediction stage, this thesis uses Neural Network (NN)-based approach to search for all anchor boxes and select the top-scoring ones, acting as a concept similar to fast filtering. It improves the detection speed and prioritizes the selection of potentially more meaningful anchor boxes. The selected anchor boxes are classified into candidate prediction boxes, further refined based on contour features. In the second prediction stage, the refined candidate prediction Boxes are fed into a Convolutional Neural Network (CNN) model for cancer recognition, thereby reducing False Positives (FP) and achieving more accurate prediction results.
This thesis proposes detection method achieves an efficiency of approximately 0.27 seconds per breast cancer MRI image, and the time required to capture a single image in an MRI scan is typically a few seconds. Therefore, it can be applied to real-time detection. Additionally, due to the cancer annotation format of the Duke Breast Cancer MRI Dataset, where a single annotation box is shared across the entire collection of breast cancer MRI images for each case, it can be challenging to precisely match cancer information in a specific breast cancer MRI image. Therefore, the research results are presented in three parts. The first part includes an Intersection over Union (IoU) greater than 0, with precision reaching 94.5%, recall reaching 89.5%, and an F-score of 92%. The second part includes an IoU greater than or equal to 0.5 while excluding breast cancer MRI images where the annotation box may significantly differ from the cancer information. This part consists of 197 validation set breast cancer MRI images, with precision reaching 83.7%, recall reaching 87.1%, AP50 reaching 77.7%, and an F-score of 85.4%. The third part selects data with IoU greater than or equal to 0.5 and predicts with YOLO (You Only Look Once version 7). The training dataset has 2365 breast cancer MRI images, and the validation dataset has 970 breast cancer MRI images, with precision reaching 61.12%, recall reaching 75%, AP50 reaching 72.8%, and an F-score reaching 67.35%.
From the research results, the proposed method is superior to YOLOv7 and can detect most cancer locations. Although the AP50 score needs improvement, the proposed method will serve as a new starting point for the automated detection of breast cancer in MRI.

目錄
摘要 i
ABSTRACT iii
誌謝 v
目錄 vi
表目錄 viii
圖目錄 ix
一、緒論 1
1.1 研究動機 1
1.2 文獻回顧 1
1.3 論文架構 3
二、相關研究 4
2.1 卷積神經網路 4
2.1.1 R-CNN 5
2.1.2 Fast R-CNN 5
2.1.3 Faster R-CNN 6
2.1.4 Mask R-CNN 6
2.1.5 YOLO 6
2.2 模型評估指標 7
2.2.1 混淆矩陣 7
2.2.2 IoU 8
2.2.3 AP50 8
2.3 影像處理 9
三、研究方法 11
3.1 資料集 11
3.2 乳癌 MRI 影像自動檢測流程 13
3.2.1 影像前處理 13
3.2.2 第一次預測 17
3.2.3 分類錨框成預測框 20
3.2.4 修正預測框 21
3.2.5 第二次預測 23
3.2.6 檢測結果輸出 25
3.2.7 與 YOLO 比較 26
四、研究成果 28
五、結論與未來展望 31
參考文獻 32
附錄A 35
附錄B 43
 
表 目 錄
表2-1 混淆矩陣 7
表3-1 乳房方向判斷模型 13
表3-2 第一次預測模型 17
表3-3 第一次預測(扁平化)模型 17
表3-4 第一次預測模型比較結果 19
表3-5 LeNet-5 模型 23
表3-6 第二次預測模型 23
表4-1 實驗結果 30
 
圖 目 錄
圖2-1 交集與聯集的示意圖 8
圖2-2 乳癌 MRI 影像經各種影像處理後示意圖 10
圖3-1 乳癌 MRI 影像自動檢測流程圖 11
圖3-2 乳癌 MRI 影像 T1 序列 13
圖3-3 乳癌 MRI 影像 T2 序列 13
圖3-4 乳癌 MRI 影像原始影像 14
圖3-5 乳癌 MRI 影像特徵圖 14
圖3-6 乳房方向判斷模型收斂曲線圖 14
圖3-7 乳癌 MRI 影像前處理示意圖 17
圖3-8 乳癌 MRI 影像前處理流程圖 17
圖3-9 第一次預測收斂曲線圖 18
圖3-10 第一次預測(扁平化)收斂曲線圖 19
圖3-11 分類錨框成預測框流程圖 20
圖3-12 乳癌 MRI 影像(未分類預測框) 21
圖3-13 乳癌 MRI 影像(分類預測框後) 21
圖3-14 修正預測框流程圖 22
圖3-15 乳癌 MRI 影像(未修正預測框) 22
圖3-16 乳癌 MRI 影像(修正預測框後) 22
圖3-17 第二次預測收斂曲線圖 24
圖3-18 檢測結果輸出流程圖 25
圖3-19 用於 YOLO 之前處理影像圖 26
圖3-20 YOLO 訓練收斂曲線圖 27
圖3-21 與 YOLO 比較流程圖 27
圖4-1 註釋框大小不合適 28
圖4-2 註釋框以外可能有腫瘤 28
圖4-3 註釋框內為空洞化之腫瘤 29
圖4-4 過於緻密之乳腺 29
圖4-5 保留註釋框大小不合適的資料，IoU 大於 0 的 PR curce 圖 30
圖4-6 排除註釋框大小不合適的資料，IoU 大於等於 0.5的 PR curve 圖 30
圖4-7 預測框與註釋框 IoU 大於等於 0.5 ，YOLO 預測的 PR curve 圖 30

[1]Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K. and Yuille, A. L., 2017. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE transactions on pattern analysis and machine intelligence, Vol. 40, no. 4, pp. 834-848.
[2]Chougrad, H., Zouaki, H. and Alheyane, O., 2018. Deep convolutional neural networks for breast cancer screening. Computer methods and programs in biomedicine, Vol. 157, pp. 19-30.
[3]Everingham, M., Van Gool, L., Williams, C. K., Winn, J. and Zisserman, A., 2010. The pascal visual object classes (VOC) challenge. International journal of computer vision, Vol. 88, pp. 303-338.
[4]Girshick, R., 2015. Fast R-CNN. Proceedings of the IEEE international conference on computer vision, pp. 1440-1448.
[5]Girshick, R., Donahue, J., Darrell, T. and Malik, J., 2014. Rich feature hierarchies for accurate object detection and semantic segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 580-587.
[6]Gu, Y., Xu, W., Lin, B., An, X., Tian, J., Ran, H., Ren, W., Chang, C., Yuan, J. and Kang, C., 2022. Deep learning based on ultrasound images assists breast lesion diagnosis in china: A multicenter diagnostic study. Insights into Imaging, Vol. 13, no. 1, Article 124.
[7]He, K., Gkioxari, G., Dollár, P. and Girshick, R., 2017. Mask R-CNN. Proceedings of the IEEE international conference on computer vision, pp. 2961-2969.
[8]He, K., Zhang, X., Ren, S. and Sun, J., 2016. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778.
[9]Hinton, G. E. and Salakhutdinov, R. R., 2006. Reducing the dimensionality of data with neural networks. science, Vol. 313, no. 5786, pp. 504-507.
[10]Jabeen, K., Khan, M. A., Alhaisoni, M., Tariq, U., Zhang, Y.-D., Hamza, A., Mickus, A. and Damaševičius, R., 2022. Breast cancer classification from ultrasound images using probability-based optimal deep learning feature fusion. Sensors, Vol. 22, no. 3, Article 807.
[11]Jaderberg, M., Simonyan, K., Zisserman, A. and Kavukcuoglu, K., 2015. Spatial transformer networks. Proceedings of the 28th International Conference on Neural Information Processing Systems, Montreal, Canada, Vol. 2, pp. 2017–2025.
[12]Jia, M., Lin, X., Zhou, X., Yan, H., Chen, Y., Liu, P., Bao, L., Li, A., Basu, P. and Qiao, Y., 2020. Diagnostic performance of automated breast ultrasound and handheld ultrasound in women with dense breasts. Breast Cancer Research and Treatment, Vol. 181, pp. 589-597.
[13]Kingma, D. P. and Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.  
[14]Krizhevsky, A., Sutskever, I. and Hinton, G. E., 2017. Imagenet classification with deep convolutional neural networks. Communications of the ACM, Vol. 60, no. 6, pp. 84-90.
[15]LeCun, Y., Bottou, L., Bengio, Y. and Haffner, P., 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE, Vol. 86, no. 11, pp. 2278-2324.
[16]Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y. and Berg, A. C., 2016. SSD: Single shot multibox detector. Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pp. 21-37.
[17]Long, J., Shelhamer, E. and Darrell, T., 2015. Fully convolutional networks for semantic segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431-3440.
[18]Lu, W., Wang, Z., He, Y., Yu, H., Xiong, N. and Wei, J., 2019. Breast cancer detection based on merging four modes MRI using convolutional neural networks. ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1035-1039.
[19]Neubeck, A. and Van Gool, L., 2006. Efficient non-maximum suppression. 18th international conference on pattern recognition (ICPR'06), Vol. 3, pp. 850-855.
[20]Niu, L., Bao, L., Zhu, L., Tan, Y., Xu, X., Shan, Y., Liu, J., Zhu, Q., Jiang, C. and Shen, Y., 2019. Diagnostic performance of automated breast ultrasound in differentiating benign and malignant breast masses in asymptomatic women: A comparison study with handheld ultrasound. Journal of Ultrasound in Medicine, Vol. 38, no. 11, pp. 2871-2880.
[21]Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics, Vol. 9, no. 1, pp. 62-66.
[22]Raaj, R. S., 2023. Breast cancer detection and diagnosis using hybrid deep learning architecture. Biomedical Signal Processing and Control, Vol. 82, Article 104558.
[23]Redmon, J. and Farhadi, A., 2018. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767.
[24]Ren, S., He, K., Girshick, R. and Sun, J., 2015. Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28.
[25]Ronneberger, O., Fischer, P. and Brox, T., 2015. U-Net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pp. 234-241.
[26]Shoshan, Y., Bakalo, R., Gilboa-Solomon, F., Ratner, V., Barkan, E., Ozery-Flato, M., Amit, M., Khapun, D., Ambinder, E. B. and Oluyemi, E. T., 2022. Artificial intelligence for reducing workload in breast cancer screening with digital breast tomosynthesis. Radiology, Vol. 303, no. 1, pp. 69-77.
[27]Simonyan, K. and Zisserman, A., 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.  
[28]Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V. and Rabinovich, A., 2015. Going deeper with convolutions. Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1-9.
[29]Uijlings, J. R., Van De Sande, K. E., Gevers, T. and Smeulders, A. W., 2013. Selective search for object recognition. International journal of computer vision, Vol. 104, pp. 154-171.
[30]Wang, C.-Y., Bochkovskiy, A. and Liao, H.-Y. M., 2023. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7464-7475.
[31]Weyrich, N. and Warhola, G. T., 1998. Wavelet shrinkage and generalized cross validation for image denoising. IEEE Transactions on Image Processing, Vol. 7, no. 1, pp. 82-90.
[32]Yuen, H., Princen, J., Illingworth, J. and Kittler, J., 1990. Comparative study of hough transform methods for circle finding. Image and vision computing, Vol. 8, no. 1, pp. 71-77.
[33]Yurttakal, A. H., Erbay, H., İkizceli, T. and Karaçavuş, S., 2020. Detection of breast cancer via deep convolution neural networks using MRI images. Multimedia Tools and Applications, Vol. 79, pp. 15555-15573.
[34]Zhang, Y., Chen, J.-H., Chang, K.-T., Park, V. Y., Kim, M. J., Chan, S., Chang, P., Chow, D., Luk, A. and Kwong, T., 2019. Automatic breast and fibroglandular tissue segmentation in breast MRI using deep learning by a fully-convolutional residual neural network U-Net. Academic radiology, Vol. 26, no. 11, pp. 1526-1535.