基于皮肤镜的黑色素瘤识别算法研究

发布时间：2018-07-28 21:06

【摘要】：黑色素瘤是一种严重威胁人类生命健康的癌症。其发病率和死亡率近年来呈逐渐增长趋势。目前的医疗水平针对转移后的黑色素瘤还缺乏有效的治疗手段。对黑色素瘤最好的治疗方式,仍是通过早期诊断并局部手术进行根除。通常,早期诊断主要依靠医生肉眼观察和组织病理学活检。然而组织病理学活检会对患者造成不必要的创伤,且肉眼观察存在极强的主观性,因此,基于皮肤镜的黑色素瘤识别算法研究是医学界的迫切需要。针对具有形态各异、颜色多变、大小不一以及噪声多样特点的皮肤镜图像处理,本文主要在图像噪声去除,皮损区域分割,皮损特征提取以及其识别分类算法上进行了研究和探索,论文主要内容有:1.针对噪声多样的皮肤镜图像,本文研究了皮肤镜图像的噪声去除算法与其分割算法。噪声去除包括有黑框、气泡等人为噪音以及毛发等固有噪音。提出的一套算法不仅能够很好的消除噪音,而且没有引入大量计算。分割算法提出了基于马尔科夫随机场融合框架的融合分割算法,提高了单个分割算法的鲁棒性;2.研究并探索了皮肤镜图像特征提取算法。基于特征设计方法提取皮肤镜图像低级特征,主要包括有形状,颜色以及纹理特征;基于特征学习方法提取中高级特征。对于特征学习方法提出了基于SIFT算子的稀疏性编码表示特征,针对皮损区域结构特点,改进池化算法,提出了环形等面积池化算法;3.研究了皮肤镜图像识别分类算法。针对提取得到的低级特征以及中高级特征,分别进行分类器设计。为更合理有效利用两类特征,对其进行分类前融合以及分类后结果融合即协同训练方法。最后利用多分类器投票机制得到最终的识别分类结果;4.基于特异性、敏感性和准确率评价准则以及受试者操作曲线(ROC)对两种特征学习得到的分类器以及融合分类器结果进行了评价。实验结果显示,融合分类器得到的结果相比单一特征识别结果的特异性、敏感性以及准确率都有所提高,同时交叉验证结果表明,融合分类器的鲁棒性也有一定提高。
[Abstract]:Melanoma is a serious threat to human life and health of cancer. Its morbidity and mortality have been increasing gradually in recent years. The current level of medical treatment for metastatic melanoma still lacks effective treatment. The best way to treat melanoma is through early diagnosis and local surgery. In general, early diagnosis mainly depends on doctors naked eye observation and histopathological biopsy. However, histopathological biopsy can cause unnecessary trauma to patients, and there is a strong subjectivity in the naked eye observation. Therefore, the research of melanoma recognition algorithm based on dermoscope is an urgent need in the medical field. Aiming at the dermoscope image processing with various shapes, colors, sizes and noises, this paper mainly focuses on image noise removal, skin lesion segmentation, The feature extraction of skin lesions and its recognition and classification algorithms are studied and explored. The main content of this paper is: 1. In this paper, the noise removal algorithm and segmentation algorithm for dermoscope images with various noises are studied. Noise removal includes black frames, bubbles and other man-made noise and hair and other inherent noise. The proposed algorithm not only can eliminate noise but also does not introduce a lot of computation. A fusion segmentation algorithm based on Markov random field fusion framework is proposed, which improves the robustness of a single segmentation algorithm. The feature extraction algorithm of dermoscope image is studied and explored. Feature based design is used to extract low-level features of dermatoscopic images, including shape, color and texture features, and feature-based learning is used to extract middle and high level features. For the feature learning method, the sparse coding feature based on SIFT operator is proposed. According to the structural characteristics of the skin lesion region, the algorithm of pool is improved, and the algorithm of ring equal area pool is proposed. The classification algorithm of dermoscope image recognition is studied. The classifier is designed for the low-level feature and middle-high-level feature. In order to make more reasonable and effective use of the two kinds of features, the pre-classification fusion and the post-classification result fusion are carried out, which is the cooperative training method. Finally, the final recognition and classification results are obtained by using the multi-classifier voting mechanism. Based on the criteria of specificity, sensitivity, accuracy and (ROC), the classifier and fusion classifier are evaluated. Experimental results show that the proposed fusion classifier is more specific, sensitive and accurate than that of single feature recognition, and the cross-validation results show that the robustness of the fusion classifier is also improved.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：R739.5;TP391.41

【相似文献】