子宮病症中較為常見之病症為子宮肌瘤(Myoma)，主因為子宮組織中肌肉腫瘤和纖維組織異常生長所導致。子宮肌腺症(Adenomyosis)則是易與子宮肌瘤混淆之病症。臨床醫師通常使用超音波和核磁共振(MRI)影像判斷子宮肌瘤及子宮肌腺症之大小、位置及相關資訊。超音波提供實時成像，成本較低且成像速度較快，然而超音波成像範圍較小，影像品質也會因為操作者受到影響。相對於超音波影像，MRI雖成像時間較長，但是提供更佳之影像品質及檢測範圍。因此，本研究使用MRI影像作為研究數據，以獲得更佳之預測結果。
在觀察醫學影像時會因為影像品質、醫師之經驗及主觀等因素有不同診斷結果，並且需要花費大量時間觀察影像以精準評估病灶位置分布。因此本研究希望透過AI人工智慧加速影像分析時間，並且以AI算法提供一致且客觀之結果。
MRI是一種灰度醫學成像技術，利用磁場對人體內分子排列之影響，創建各種器官和組織之對比圖像。每組MRI切片約包含30張切片，切片間隔約為5.5毫米。不同磁共振設備獲得之MRI圖像亮度和灰度值可能會有所不同。在本研究中，放射科醫生首先對子宮、子宮肌瘤和子宮腺肌症數據集進行標註。其中資料集以641個子宮標註影像、762個子宮肌瘤標註影像和149個子宮腺肌症標註影像進行模型訓練。其中對子宮肌瘤數據集進行對比度和亮度值調整，以增強圖像識別能力並重新訓練。針對子宮預測影像採用輪廓修正演算法提升預測結果。
本研究利用深度學習模型對子宮、子宮肌瘤及子宮肌腺症進行訓練並預測其位置和大小。訓練集會先進行資料增量如放大、縮小、旋轉……，接著使用深度學習語意分割模型ESFPNet進行訓練，損失函數採用cross-entropy，評估函數則使用Dice score作為參考依據。訓練並經由演算法修正預測影像後，子宮影像預測之Dice score值達85.29%，子宮輪廓修正後達86.13%，子宮肌瘤達77.72%，子宮肌腺症達68.12%，並將預測結果提供醫師以輔助臨床診斷和治療計畫。

Uterine fibroid (myoma) is a relatively common condition, which is characterized by abnormal growth of muscular tumor with fibrotic tissue in the uterine tissue. Adenomyosis is a condition that can be easily confused with uterine fibroids. Clinicians typically use ultrasonography and magnetic resonance imaging (MRI) to assess the size, location, and other information related to uterine fibroids and adenomyosis. Ultrasound provides real-time imaging with lower costs and faster outputs for diagnosis. However, the quality and coverage of the images are relatively limited and operator-dependent. In contrast, MRI is a longer procedure and however offers better resolution and range of detection. Hence, this study utilizes MRI as the research data to achieve higher-quality predictive results.
In a medical image, the diagnosis can vary because of such factors as image quality, physician’s experience, and subjectivity. Moreover, lengthy examination of the images is required for accurate evaluation of lesion location and distribution. Therefore, this study aimed to utilize artificial intelligence (AI) to expedite the image analysis process, and consequently provide consistent and objective interpretation of the images through AI algorithms.
MRI is a grayscale medical imaging technology that uses the effects of magnetic fields on molecular alignment within the human body to create contrast images of various organs and tissues. Each set of MRI slices consists of approximately 30 slices, with a slice interval of approximately 5.5mm. The brightness and grayscale values of MRI images obtained from different magnetic resonance machines can vary. This study primarily employs a model training and background removal approach to standardize MRI images from different hospitals. In this study, the datasets for the uterus, uterine fibroids, and adenomyosis were first annotated by radiologists. A total of 641 annotated images of the uterus, 762 annotated images of uterine fibroids, and 149 annotated images of adenomyosis were used for model training. The uterine fibroid dataset was adjusted for contrast and brightness values of the predicted images to enhance image recognition, and the images were retrained. The contour correction algorithm was applied to improve the quality of the predicted images for the uterus.
Then, deep learning models were utilized to train models for the uterus, uterine fibroids, and adenomyosis to predict their respective positions and sizes. The training set underwent data augmentation, such as zooming, shrinking, and rotation, followed by training using the deep learning semantic segmentation model ESFPNet. The cross-entropy loss function was used. The Dice score was employed as the evaluation function. After training and algorithm-based prediction image correction, the Dice score for predicting uterine images reached 85.29%. After contour correction for the uterus, it reached 86.13%. For uterine fibroids and adenomyosis, the score was 77.72% and 68.76%, respectively. The results of this study will facilitate clinical diagnosis and therapeutic planning by the clinicians.

論文審定書 i
中文摘要 ii
Abstract iv
目錄(Table of Contents) vii
圖次 ix
表次 xi
第一章 緒論 1
1.1背景 1
1.2目的 2
1.3研究大綱 3
第二章 相關研究 4
2.1 深度學習 4
2.2 Vision Transformer(ViT) 4
2.2.1 編碼器與解碼器結構 5
2.2.2 多頭注意力(Multi-Head Attention) 6
2.3 語意分割(Semantic segmentation) 6
2.4 ESFPNet神經網路 8
2.5 資料集 9
2.5.1數據集與醫療影像 9
2.5.2遷移學習 11
2.5.3資料增量 12
第三章 研究流程與方法 14
3.1 訓練流程 14
3.2 影像處理 14
3.2.1計算子宮區域亮度 14
3.2.2對比度計算與調整 15
3.3 輪廓修正演算法 16
3.4 資料增量方式 18
3.5 深度模型訓練 19
第四章 研究結果 20
4.1 子宮預測結果 20
4.2 子宮預測影像後處理後之結果比較 24
4.3 子宮肌瘤預測結果 26
4.4 子宮肌瘤對比度調整訓練及預測結果 29
4.5 子宮肌腺症訓練結果 32
第五章 結論與未來展望 36
參考文獻 37

[1] N. A. Bridges, A. Cooke, M.J. Healy, P.C. Hindmarsh, and C.G. Brook, “Growth of the uterus,” Archives of disease in childhood, 1996
[2] Uri Goldsztejn, and Arye Nehorai, “Estimating uterine activity from electrohysterogram measurements via statistical tensor decomposition,” arXiv preprint arXiv: 2209.02183, September 2022
[3] Qing Lyu, and Ge Wang, “Conversion Between CT and MRI Images Using Diffusion and Score-Matching Models,” arXiv preprint arXiv: 2209.12104, September 2022
[4] Lydia Garcia, and Keith Isaacson, Adenomyosis: review of the literature, Journal of minimally invasive gynecology, pp. 428-437, 2011
[5] Ryan Prescott Adams, Hanna M. Wallach, and Zoubin Ghahramani, “Learning the Structure of Deep Sparse Graphical Models,” arXiv preprint arXiv: 1001.0160, 2010
[6] Juergen Schmidhuber, “Deep Learning in Neural Networks: An Overview,” arXiv preprint arXiv: 1404.7828, 2014
[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby, “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv preprint arXiv: 2010.11929, 2021
[8] Amirsina Torfi, Rouzbeh A. Shirvani, Yaser Keneshloo, Nader Tavaf, and Edward A. Fox, “Natural Language Processing Advancements By Deep Learning: A Survey,” arXiv preprint arXiv: 2003.01200, 2021
[9] Keiron O'Shea, and Ryan Nash, “An Introduction to Convolutional Neural Networks,” arXiv preprint arXiv: 1511.08458, 2015
[10] Sherstinsky Alex, “Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network,” arXiv preprint arXiv: 1808.03314, August 2018
[11] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin, “Attention Is All You Need,” arXiv preprint arXiv: 1706.03762, June 2017
[12] Keiron O'Shea , and Ryan Nash, “On Efficient Real-Time Semantic Segmentation: A Survey,” arXiv preprint arXiv: 2206.08605, June 2015
[13] Qi Chang, Danish Ahmad, Jennifer Toth, Rebecca Bascom, and William E. Higgins, “ESFPNet: efficient deep learning architecture for real-time lesion segmentation in autofluorescence bronchoscopic video,” arXiv preprint arXiv: 2207.07759, July 2022
[14] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo, “SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers,” arXiv preprint arXiv: 2105.15203, May 2021
[15] Vincent Dumoulin, and Francesco Visin, “A guide to convolution arithmetic for deep learning,” arXiv preprint arXiv: 1603.07285, March 2016
[16] Suorong Yang, Weikang Xiao, Mengcheng Zhang, Suhan Guo, Jian Zhao, and Furao Shen, “Image Data Augmentation for Deep Learning: A Survey,” arXiv preprint arXiv: 2204.08610, April 2022