隨著計算機(jī)視覺技術(shù)的進(jìn)步,醫(yī)學(xué)圖像處理已被廣泛用于醫(yī)療診斷。近年來,為了提高癌癥早期檢測準(zhǔn)確率并改善治療的效果,特別是在腦癌、肺癌、乳腺癌等腫瘤結(jié)節(jié)的診斷。超聲圖像的分辨率較低,并且圖像中顯示腫瘤的區(qū)域通常是模糊的,如邊緣模糊,形狀不規(guī)則。甲狀腺結(jié)節(jié)是人體內(nèi)分泌系統(tǒng)常最見的疾病,自動實現(xiàn)結(jié)節(jié)的良惡性分類可以輔助醫(yī)生進(jìn)行相關(guān)疾病的診斷。本質(zhì)上,大多數(shù)甲狀腺結(jié)節(jié)是良性的,只有不到5%是惡性的。大多數(shù)現(xiàn)有技術(shù)都有一些局限性,因為這些技術(shù)是在有限的數(shù)據(jù)集下進(jìn)行檢查和評估并且沒有實現(xiàn)自動分類。因此,為了更加準(zhǔn)確的實現(xiàn)甲狀腺結(jié)節(jié)良惡性自動分類,我們將卷積神經(jīng)網(wǎng)絡(luò)應(yīng)用于甲狀腺結(jié)節(jié)超聲圖像的分類。為了提高檢測的準(zhǔn)確性,我們使用了深度卷積神經(jīng)網(wǎng)絡(luò)VGG-16提取結(jié)節(jié)區(qū)域的特征用于分類。在本研究中,VGG-16被用于甲狀腺結(jié)節(jié)的分類。我們在公共數(shù)據(jù)集和本地數(shù)據(jù)集上訓(xùn)練和測試了模型。該模型可以較為快速、可靠地對甲狀腺癌結(jié)節(jié)的良惡性進(jìn)行分類,在醫(yī)學(xué)領(lǐng)域具有一定的應(yīng)用價值。 

【文章來源】：華南理工大學(xué)廣東省 211工程院校 985工程院校 教育部直屬院校

【文章頁數(shù)】：67 頁

【學(xué)位級別】：碩士

【文章目錄】：
摘要
ABSTRACT
CHAPTER 1 INTRODUCTION
    1.1 Medical Ultrasound Therapy
    1.2 Medical Ultrasound in Clinical Medicine
        1.2.1 Therapeutic application
    1.3 Thyroid Disease
        1.3.1 Thyroid Nodule & Cancer
        1.3.2 Types of Thyroid Cancer
    1.4 Thyroid Disorder Diagnosis
        1.4.1 Blood Test
        1.4.2 Imaging test
        1.4.3 Biopsy
CHAPTER 2 MACHINE LEARNING ALGORITHMS
    2.1 Approaches of Machine
        2.1.1 Supervised machine learning
        2.1.2 Unsupervised machine learning
        2.1.3 Semi-supervised learning
    2.2 Artificial Neural Networks (ANNs)
        2.2.1 Feed forward Neural Networks
        2.2.2 Recurrent Networks
    2.3 Neural Networks
        2.3.1 Convolutional Neural Networks
    2.4 CNN Architecture
        2.4.1 LeNet-5 (1998)
        2.4.2 AlexNet (2012)
        2.4.3 ZFNet (2013)
        2.4.4 Google Net/Inception (2014)
        2.4.5 VGGNet (2014)
        2.4.6 ResNet (2015)
    2.5 Recurrent Neural Networks
    2.6 Long Short-Term Memory
    2.7 Radial Basis Function Network
    2.8 Capsule Neural Network
    2.9 Bayesian Networks
    2.10 Support Vector Machines
CHAPTER 3 VGG 16 MODEL
    3.1 VGG-16
    3.2 Architecture
    3.3 Components of VGG 16
        3.3.1 Convolution layer
        3.3.2 Re LU layer
        3.3.3 Pooling layer
        3.3.4 Batch normalization layer
        3.3.5 Dropout
        3.3.6 Soft Max, Loss and Regularization
        3.3.7 Optimization
    3.4 Implementation Details
    3.5 Algorithm
        3.5.1 Forward Propagation
        3.5.2 Parameters Initialization
        3.5.3 Activation Functions
        3.5.4 Rectified Linear Unit (Re LU)
        3.5.5 Leaky Rectified Linear Unit
        3.5.6 Feed Forward
        3.5.7 Cost
        3.5.8 Back-Propagation
CHAPTER 4 EXPERIMENTAL DATASET & RESULTS
    4.1 Dataset
    4.2 Image Pre-processing
    4.3 Data Augmentation
    4.4 Software and Hardware
    4.5 Experimental Procedures
        4.5.1 Splitting Dataset:
        4.5.2 Load The Dataset
        4.5.3 Train the model
    4.6 Results
    4.7 Discussion
CHAPTER 5 CONCLUSION
    Future work
References
Acknowledgement
附件



本文編號：3521612