【摘要】：黑色素瘤是在臨床中經(jīng)常遇到的一種惡性皮膚腫瘤,同時也是世界上增長最快的癌癥之一。通常皮膚科醫(yī)生通過肉眼觀察和組織病理學(xué)活檢的方法來對黑色素瘤進(jìn)行早期的篩查和確診。盲目進(jìn)行活檢常常會對患者造成經(jīng)濟(jì)上的壓力和身體上不必要的創(chuàng)傷。因此,非創(chuàng)傷性的黑色素瘤自動診斷技術(shù)成為了醫(yī)學(xué)界急需解決的問題。皮膚鏡圖像形態(tài)模式識別一直是分辨良性腫瘤和惡性素瘤的一項具有挑戰(zhàn)性的任務(wù)。因此針對黑色素瘤皮膚鏡圖像的多種形態(tài)模式,本文對基于多標(biāo)簽學(xué)習(xí)的惡性黑色素瘤形態(tài)模式識別進(jìn)行了深入的研究。創(chuàng)新點主要包括以下部分:1.深入研究了黑色素瘤的分割與特征提取方法。為了更好的提取皮膚鏡圖像的特征,提出一種基于區(qū)域一致性的融合算法來對圖像進(jìn)行分割,將多個分割算法結(jié)果進(jìn)行融合,依據(jù)區(qū)域大小、灰度值、紋理的一致性原則移除與融合結(jié)果相矛盾的子區(qū)域,從而得到最終的分割結(jié)果。對分割結(jié)果分別提取皮損內(nèi)區(qū)域、皮損區(qū)域和皮損外區(qū)域的顏色特征、形狀特征和紋理特征。2.提出了基于手工特征提取的多標(biāo)簽分類在黑色素瘤模式識別中的應(yīng)用。對黑色素瘤的形態(tài)模式種類進(jìn)行了深入研究分析,可明確定義的全局形態(tài)特征主要有八種,其中涉及七種基本模式和一個多成分模式。這七種基本模式包括:網(wǎng)狀模式、球狀模式、鵝卵石模式、星爆模式、平行模式、腔洞模式、均勻模式。通過黑色素瘤的七種基本模式來建立多標(biāo)簽分類模型,以達(dá)到自動識別皮膚鏡圖像中所包含的模式類別的目的。使用Binary Reference算法和ML-kNN算法對黑色素瘤特征向量進(jìn)行多標(biāo)簽分類,對兩種算法的多標(biāo)簽分類結(jié)果對比分析發(fā)現(xiàn)ML-kNN算法對黑色素瘤的多種形態(tài)模式的識別相比于Binary Reference算法具有更好的效果。3.提出了基于特征學(xué)習(xí)的卷積神經(jīng)網(wǎng)絡(luò)多標(biāo)簽分類在黑色素瘤模式識別中的應(yīng)用。在深度學(xué)習(xí)框架的基礎(chǔ)上提出了一種改進(jìn)的方法來實現(xiàn)多標(biāo)簽分類,將圖像數(shù)據(jù)和多標(biāo)簽數(shù)據(jù)分別作為網(wǎng)絡(luò)的輸入層,然后通過在網(wǎng)絡(luò)結(jié)構(gòu)添加Slice層達(dá)到多標(biāo)簽分類的目的,最終得到多標(biāo)簽分類模型。在利用卷積神經(jīng)網(wǎng)絡(luò)對黑色素瘤的形態(tài)模式進(jìn)行特征自動學(xué)習(xí)的實驗中,實驗結(jié)果表明本文提出的利用卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行多標(biāo)簽分類效果比基于手工特征的多標(biāo)簽分類具有顯著的提升。
[Abstract]:Melanoma is a malignant skin tumor often encountered in clinical practice, and it is also one of the fastest growing cancers in the world. Dermatologists usually screen and diagnose melanoma early through naked eye observation and histopathological biopsy. Blind biopsies often cause financial stress and unnecessary physical trauma to patients. Therefore, non-traumatic automatic diagnosis of melanoma has become an urgent problem in medical field. Pattern recognition of dermatoscopic images has been a challenging task in distinguishing benign and malignant tumors. Therefore, for the multiple morphologic patterns of melanoma dermoscope images, this paper studies the morphological pattern recognition of malignant melanoma based on multi-label learning. Innovations include the following parts: 1. The segmentation and feature extraction of melanoma were studied. In order to extract the features of skin mirror image better, a fusion algorithm based on region consistency is proposed to segment the image. The results of multiple segmentation algorithms are fused according to the region size and gray value. The consistency principle of texture removes the subregions which are contradictory to the fusion results, and the final segmentation results are obtained. The color features, shape features and texture features of the inner region, the lesion region and the outer area of the skin lesions were extracted from the segmentation results. The application of multi-label classification based on manual feature extraction in melanoma pattern recognition is proposed. The morphological patterns of melanoma were studied and analyzed. There are eight kinds of global morphological features which can be clearly defined, including seven basic patterns and one multi-component pattern. The seven basic models include reticular model, spherical model, cobblestone model, starburst model, parallel mode, cavity model, and uniform mode. Through seven basic patterns of melanoma, a multi-label classification model is established to automatically identify the pattern categories contained in the dermatoscopic image. Binary Reference algorithm and ML-kNN algorithm are used to classify melanoma feature vectors. The comparison and analysis of the multi-label classification results of the two algorithms show that the ML-kNN algorithm has a better effect than the Binary Reference algorithm in the recognition of multiple morphologic patterns of melanoma. This paper presents the application of convolution neural network multi-label classification based on feature learning in melanoma pattern recognition. Based on the deep learning framework, an improved method is proposed to realize multi-label classification. Image data and multi-label data are used as the input layer of the network, and then the multi-label classification is achieved by adding the Slice layer to the network structure. Finally, multi-label classification model is obtained. In the experiment of using convolutional neural network to study the morphologic pattern of melanoma, The experimental results show that the effectiveness of multi-label classification based on convolution neural network is significantly improved than that of multi-label classification based on manual features.
【學(xué)位授予單位】：電子科技大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：R739.5;TP391.41

【參考文獻(xiàn)】

相關(guān)期刊論文 前2條

1 孟如松;孟曉;姜志國;謝鳳英;劉瑋;羅衛(wèi);郭廣進(jìn);蔡瑞康;;基于國人皮膚鏡黑素細(xì)胞腫瘤圖像的智能化分類與識別研究[J];中國體視學(xué)與圖像分析;2012年03期

2 孟如松;趙廣;;皮膚鏡圖像分析技術(shù)的基礎(chǔ)與臨床應(yīng)用[J];臨床皮膚科雜志;2008年04期

，

本文編號：2187624