乳腺癌是最常被發(fā)現(xiàn)的癌癥之一,它威脅到女性的健康和生命。早期準(zhǔn)確的疾病診斷在癌癥治療中起著重要作用。臨床研究表明,如果在早期階段檢測(cè)到癌癥,它可以相對(duì)容易地治愈而不會(huì)對(duì)患者造成太大傷害。超聲成像是檢測(cè)乳腺癌的一種方法。醫(yī)療超聲波使用人體無(wú)法聽(tīng)到的高頻聲波（>20,000 Hz）,將脈沖發(fā)送到人體組織中并以不同的屬性反射回來(lái)被記錄并顯示為圖像,是一種觀察人體內(nèi)腫瘤和其他異常等疾病的便捷工具。本文的主要任務(wù)為減少癌癥診斷中的錯(cuò)誤,使醫(yī)用計(jì)算機(jī)可以自動(dòng)發(fā)現(xiàn)腫瘤區(qū)域及周圍組織結(jié)構(gòu)。實(shí)現(xiàn)此任務(wù)的一種方法是應(yīng)用圖像語(yǔ)義分割。圖像語(yǔ)義分割不僅可以計(jì)算對(duì)象位置,還可以確定其所屬類別。分類模型用于計(jì)算每個(gè)像素的概率分布,以優(yōu)化結(jié)果進(jìn)行準(zhǔn)確分割。本文首先總結(jié)了國(guó)內(nèi)外圖像分割的研究現(xiàn)狀,然后提出了圖像語(yǔ)義分割方法的思路。這項(xiàng)研究的主要思想是應(yīng)用全卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行特征提取。最初,FCN使用訓(xùn)練樣本和單熱圖像進(jìn)行訓(xùn)練,然后在提供另一組圖像之后,可以生成分割結(jié)果。全卷積網(wǎng)絡(luò)的架構(gòu)主要采用U-net及其他兩種變體,將通過(guò)實(shí)驗(yàn)分別測(cè)試三種架構(gòu)的性能。這兩個(gè)稱為Dual U-net與Tight U-net的變體基... 

【文章來(lái)源】：哈爾濱工業(yè)大學(xué)黑龍江省 211工程院校 985工程院校

【文章頁(yè)數(shù)】：130 頁(yè)

【學(xué)位級(jí)別】：碩士

【文章目錄】：
摘要
Abstract (In English)
Chapter 1 Introduction
    1.1 Background
    1.2 Research inside and outside of country and analysis
        1.2.1 Overview of non-semantic segmentation methods
        1.2.2 Overview of semantic segmentation methods
    1.3 Problems of segmentation methods
    1.4 Research summary
    1.5 Organization of the thesis
Chapter 2 Fundamentals of Convolutional Neural Networks
    2.1 Image Augmentation
    2.2 Confusion matrix and Jaccard index
    2.3 Architecture of Convolution neural network
        2.3.1 Convolution layer
        2.3.2 Pooling Layer
        2.3.3 Fully Connected Layer
        2.3.4 Activation functions
    2.4 Original U-net Architecture
        2.4.1 Choosing the architecture of Fully Convolutional Network
        2.4.2 The description of the original architecture
        2.4.3 Re LU
    2.5 Training the network
        2.5.1 Loss function
        2.5.2 Optimization algorithm
    2.6 Wavelet Image Transformation
        2.6.1 DWT in two dimensions
        2.6.2 The application of DWT in the research
Chapter 3 Segmentation with Dual U-net
    3.1 Main Contribution
        3.1.1 U-net and Its Limitations
        3.1.2 Dual U-net
    3.2 Application of Original U-net and Dual U-net
    3.3 Training
    3.4 Experiments and results
        3.4.1 Two-class classification
        3.4.2 Multi-class classification
    3.5 Brief Summary
Chapter 4 Segmentation with Tight U-net
    4.1 Architecture
    4.2 Application of Tight U-net
    4.3 Training
    4.4 Experiments and results
        4.4.1 Two-class classification
        4.4.2 Multi-class classification
    4.5 Brief Summary
Chapter 5 Optimization of segmentation results of U-net with CRF
    5.1 Mean-field approach
    5.2 CRF application
    5.3 Experiments and results
        5.3.1 Two-class classification
        5.3.2 Multi-class classification
    5.4 Brief Summary
Conclusion (In English)
結(jié)論
References
Acknowledgement
Resume



本文編號(hào)：3182533