【摘要】：上皮和間質(zhì)組織是乳腺組織病理圖像中最基本的兩種組織,約80%的乳腺腫瘤起源于乳腺上皮組織.為了構(gòu)建基于乳腺組織病理圖像分析的計(jì)算機(jī)輔助診斷系統(tǒng)和分析腫瘤微環(huán)境,上皮和間質(zhì)組織的自動(dòng)分割是重要的前提條件.本文構(gòu)建一種基于逐像素點(diǎn)深度卷積網(wǎng)絡(luò)(CN-PI)模型的上皮和間質(zhì)組織的自動(dòng)分割方法.1)以病理醫(yī)生標(biāo)注的兩類區(qū)域邊界附近具有類信息為標(biāo)簽的像素點(diǎn)為中心,構(gòu)建包含該像素點(diǎn)上下文信息的正方形圖像塊的訓(xùn)練集.2)以每個(gè)正方形圖像塊包含的像素的彩色灰度值作為特征,以這些圖像塊中心像素類信息為標(biāo)簽訓(xùn)練CN模型.在測(cè)試階段,在待分割的組織病理圖像上逐像素點(diǎn)地取包含每個(gè)中心像素點(diǎn)上下文信息的正方形圖像塊,并輸入到預(yù)先訓(xùn)練好的CN網(wǎng)絡(luò)模型,以預(yù)測(cè)該圖像塊中心像素點(diǎn)的類信息.3)以每個(gè)圖像塊中心像素為基礎(chǔ),逐像素地遍歷圖像中的每一個(gè)像素,將預(yù)測(cè)結(jié)果作為該圖像塊中心像素點(diǎn)類信息的預(yù)測(cè)標(biāo)簽,實(shí)現(xiàn)對(duì)整幅圖像的逐像素分割.實(shí)驗(yàn)表明,本文提出的CN-PI模型的性能比基于圖像塊分割的CN網(wǎng)絡(luò)(CN-PA)模型表現(xiàn)出了更優(yōu)越的性能.
[Abstract]:Epithelium and mesenchymal tissue are the two most basic tissues in breast histopathological images. About 80% of breast tumors originate from mammary epithelial tissue. In order to construct a computer-aided diagnosis system based on breast pathological image analysis and analyze tumor microenvironment, automatic segmentation of epithelial and interstitial tissues is an important prerequisite. In this paper, an automatic segmentation method of epithelial and interstitial tissue based on pixel by pixel depth convolution network (CN-PI) model is proposed. A training set of square image blocks containing the context information of the pixel is constructed. 2) the color gray value of the pixels contained in each square image block is used as the feature, and the central pixel class information of these blocks is used as the label to train the CN model. In the test phase, the square image blocks containing the context information of each central pixel point are selected from the histopathological image to be segmented, and input to the pre-trained CN network model. Based on the class information of predicting the central pixel of the image block. 3) based on the central pixel of each image block, every pixel in the image is traversed pixel by pixel, and the prediction result is used as the prediction label of the pixel class information in the center of the image block. The pixel-by-pixel segmentation of the whole image is realized. Experimental results show that the performance of the proposed CN-PI model is better than that of the CN neural network (CN-PA) model based on image block segmentation.
【作者單位】： 南京信息工程大學(xué)江蘇省大數(shù)據(jù)分析技術(shù)重點(diǎn)實(shí)驗(yàn)室;武漢大學(xué)中南醫(yī)院腫瘤科腫瘤生物學(xué)行為湖北省重點(diǎn)實(shí)驗(yàn)室 湖北省腫瘤醫(yī)學(xué)臨床研究中心;
【基金】：國(guó)家自然科學(xué)基金(61771249,61273259) 江蘇省“六大人才高峰”高層次人才項(xiàng)目資助計(jì)劃(2013-XXRJ-019) 江蘇省自然科學(xué)基金(BK20141482) 江蘇創(chuàng)新創(chuàng)業(yè)團(tuán)隊(duì)人才計(jì)劃(JS201526)資助~~
【分類號(hào)】：R737.9;TP391.41
，

本文編號(hào)：2266316