本文關(guān)鍵詞： 醫(yī)學(xué)圖像識別 卷積神經(jīng)網(wǎng)絡(luò) 特征學(xué)習(xí) 多任務(wù)學(xué)習(xí)　出處：《中國科學(xué)技術(shù)大學(xué)》2017年博士論文　論文類型：學(xué)位論文

【摘要】：醫(yī)學(xué)成像技術(shù)與相關(guān)圖像識別算法的快速發(fā)展反映了人們對醫(yī)學(xué)信息獲取的強(qiáng)烈需求。醫(yī)學(xué)圖像能夠提供豐富的信息,在醫(yī)學(xué)診斷中的作用日益凸顯。計(jì)算機(jī)識別算法能夠克服人工識別易受認(rèn)知能力、主觀經(jīng)驗(yàn)、疲勞程度影響的不足,有效提高識別的準(zhǔn)確率和穩(wěn)定性,減少誤診和漏診,對病情診斷、病理分析及治療方案的選取有重大意義。卷積神經(jīng)網(wǎng)絡(luò)能夠提供基于學(xué)習(xí)的特征表示,基于卷積神經(jīng)網(wǎng)絡(luò)模型的圖像分類、檢測、分割算法在醫(yī)學(xué)圖像識別上有廣泛應(yīng)用。本文的研究內(nèi)容的是基于卷積神經(jīng)網(wǎng)絡(luò)的圖像分類和語義分割算法,應(yīng)用于醫(yī)學(xué)圖像中癌變目標(biāo)識別。主要工作包括:1)基于CNN-SVM的微血管分型識別算法研究。微血管分型與癌癥發(fā)展密切相關(guān),分型識別是食道癌診斷與治療的前提。CNN由數(shù)據(jù)驅(qū)動(dòng),相比手工設(shè)計(jì)特征更加適合復(fù)雜多變的微血管圖像。本文在樣本量相對較少的情況下,設(shè)計(jì)了一個(gè)以CNN-SVM為核心模型的微血管分型識別系統(tǒng),對一系列的數(shù)據(jù)擴(kuò)增技術(shù)進(jìn)行了研究,逐步地提升系統(tǒng)對縮放、旋轉(zhuǎn)圖像預(yù)測的魯棒性;在分類器提升方面,引入SVM替換softmax增強(qiáng)了系統(tǒng)的泛化能力。對比廣泛使用手工設(shè)計(jì)特征,CNN彰顯了優(yōu)越的特征表達(dá)能力。2)基于多約束FCN的微血管分型語義分割算法研究。本文提出采用語義分割算法對微血管分型進(jìn)行識別。針對不完全標(biāo)注問題,結(jié)合人工知識,從標(biāo)注信息中挖掘出感興趣區(qū)域信息,構(gòu)建了一個(gè)基于多約束FCN的語義分割系統(tǒng)。感興趣區(qū)域標(biāo)簽包含了人工積累的經(jīng)驗(yàn),該系統(tǒng)采用多任務(wù)學(xué)習(xí)框架,利用多種標(biāo)簽提升了編碼器類間區(qū)分能力,從而提高了 FCN網(wǎng)絡(luò)的分割性能。3)基于聯(lián)合學(xué)習(xí)FCN的細(xì)胞圖像語義分割算法研究。顯微鏡下癌變細(xì)胞識別是病理檢查的主要內(nèi)容,也是癌癥確診的關(guān)鍵。對癌細(xì)胞區(qū)域進(jìn)行精細(xì)劃分十分困難。本文采用語義分割算法對癌變區(qū)域進(jìn)行識別。根據(jù)多任務(wù)學(xué)習(xí)思想,設(shè)計(jì)了分類任務(wù),提出CNN與FCN的聯(lián)合學(xué)習(xí)方法。通過對分類任務(wù)的探索,完成了模型優(yōu)化,并對額外數(shù)據(jù)集的價(jià)值進(jìn)行了驗(yàn)證。在多任務(wù)學(xué)習(xí)的框架下,通過提升分類任務(wù)性能間接地改善了分割任務(wù)的性能。綜上所述,本文從模型和數(shù)據(jù)兩個(gè)方面來提升卷積神經(jīng)網(wǎng)絡(luò)的特征表達(dá)能力。模型方面的研究工作包括設(shè)計(jì)網(wǎng)絡(luò)結(jié)構(gòu)、引入SVM、采用BN規(guī)范化和探究基于梯度下降的優(yōu)化算法;數(shù)據(jù)方面的研究工作包括探索數(shù)據(jù)集擴(kuò)增技術(shù),設(shè)計(jì)感興趣區(qū)域標(biāo)簽,利用額外有價(jià)值數(shù)據(jù)。實(shí)驗(yàn)表明這些改進(jìn)能提升識別性能。
[Abstract]:The rapid development of medical imaging technology and image recognition algorithm reflects the strong demand for medical information. Medical image can provide abundant information in medical diagnosis is playing an increasingly prominent role. The computer recognition algorithm can overcome the artificial recognition by cognitive ability, subjective experience, lack of fatigue effect, effectively improve the accuracy the recognition rate and stability, reduce misdiagnosis and missed diagnosis, the diagnosis, pathological analysis and selection of treatment plan is of great significance. The convolution neural network can provide learning based on the characteristics of representation, image classification, based on the model of convolutional neural network detection, segmentation algorithm has been widely used in medical image recognition on the research contents of this paper. The image classification and semantic segmentation algorithm based on convolutional neural network is applied to target recognition, canceration of the medical image. The main work includes: 1) base Study on the recognition algorithm is divided in CNN-SVM. Microvessel microvessel typing and cancer is closely related to the development, type recognition is the premise of.CNN for the diagnosis and treatment of esophageal cancer by data driven, manual design features compared to more suitable for micro vascular image is complex. This paper is relatively less in the sample volume, design a with CNN-SVM as the core model of the micro vascular pattern recognition system, amplification of a series of data, and gradually improve the system robustness zoom, rotate the image prediction; in the classifier upgrade, the introduction of SVM to replace softmax to enhance the generalization ability of the system. Compared with widely used manual design features, highlighting the CNN the superior feature representation ability of.2) microvascular multi constrained FCN type segmentation algorithm based on semantic study. This paper adopts semantic segmentation algorithm for micro vascular pattern recognition. At the end The annotation problem, combined with artificial knowledge, from the annotation information mining region of interest information, constructs a semantic segmentation system based on multi constrained FCN. Region of interest label contains the artificial experience, the system adopts multi task learning framework, the label to enhance the ability to distinguish between types of encoder using a variety, and to improve the segmentation performance of.3 FCN network) segmentation algorithm based on semantic association learning cell image based on FCN. Identification of cancerous cells under the microscope is the main content of pathological examination, is also a key cancer diagnosis. It is very difficult to fine division of cancer cell region. This paper uses the semantic segmentation algorithm is used to identify the cancerous area. According to multi task the thought of learning, design the classification task, proposed the joint learning method of CNN and FCN. Through the exploration of the classification task, complete the optimization model, and the additional data set The value is verified. In the framework of multi task learning, by improving the classification task performance indirectly improves the performance of the segmentation task. To sum up, this paper to improve the characteristics of convolutional neural network from two aspects of model and data expression ability. Research model including design of network structure, the introduction of SVM, using the BN specification and explore the optimization algorithm based on gradient descent; research data include exploration data set amplification technology, design ROI label, using additional valuable data. Experiments show that these improvements can improve the recognition performance.

【學(xué)位授予單位】：中國科學(xué)技術(shù)大學(xué)
【學(xué)位級別】：博士
【學(xué)位授予年份】：2017
【分類號】：TP391.41;TP183

【參考文獻(xiàn)】

相關(guān)期刊論文 前2條

1 莊福振;羅平;何清;史忠植;;遷移學(xué)習(xí)研究進(jìn)展[J];軟件學(xué)報(bào);2015年01期

2 劉志剛,李德仁,秦前清,史文中;支持向量機(jī)在多類分類問題中的推廣[J];計(jì)算機(jī)工程與應(yīng)用;2004年07期

相關(guān)博士學(xué)位論文 前1條

1 馮慧;上消化道早癌篩查、診斷及其相關(guān)技術(shù)的探索性研究[D];安徽醫(yī)科大學(xué);2016年

，

本文編號：1476901