论文部分内容阅读
胰腺癌的诊断非常重要,而细胞抹片显微图像的病理分析是其诊断的主要手段。图像的准确自动分割和分类是病理分析的重要环节,因此本文提出了一种新的胰腺细胞抹片显微图像自动分割与分类算法。在分割方面,首先采用多特征Mean-shift聚类算法(MFMS)定位细胞核区域;接着采用弹性数学形态学结合角点检测的去粘连模型(CSM)对粘连重叠细胞核进行去粘连处理,实现了分割的准确性和鲁棒性。在分类方面,首先针对分割的细胞核提取了4个形状特征和138个不同颜色空间的纹理特征;然后结合支持向量机(SVM)和链式遗传算法(CAGA)实现封装式特征选择;最后将优选特征送入SVM进行分类,完成了胰腺细胞抹片显微图像的分类识别。本文采用了15幅图像一共461个细胞核进行测试。实验结果显示,本文算法可以实现不同类型的胰腺细胞抹片显微图像的自动分割与准确分类。就分割来说,本文算法可获得较高的正确率(93.46%±7.24%);就正常和癌变细胞的分类来说,本文算法可获得较高的分类正确率(96.55%±0.99%)、灵敏度(96.10%±3.08%)和特异度(96.80%±1.48%)。
The diagnosis of pancreatic cancer is very important, and pathological analysis of the cell smear microscopy is the main means of diagnosis. Accurate automatic segmentation and classification of images is an important part of pathological analysis. Therefore, this paper presents a new automatic segmentation and classification algorithm of pancreatic cell smear microscopic images. In the aspect of segmentation, first, the multi-feature Mean-shift clustering algorithm (MFMS) was used to locate the nucleus region; then the dendritic adhesion model (CSM) was used to debond the overlapping nuclei by elastic mathematical morphology and corner detection, Accuracy and robustness. In terms of classification, four shape features and 138 texture features in different color spaces are extracted firstly for the segmented nuclei. Then, feature selection of package features is implemented by combining Support Vector Machine (SVM) and Chain Genetic Algorithm (CAGA). Finally, Feature into the SVM for classification, completed the pancreatic cell smear microscopic image classification and identification. This article uses a total of 46 images of 15 nuclei for testing. Experimental results show that the proposed algorithm can automatically segment and classify different types of pancreatic cell smears microscopic images. In terms of segmentation, the proposed algorithm can achieve a high accuracy rate (93.46% ± 7.24%). For the classification of normal and cancerous cells, our algorithm can obtain a higher classification accuracy (96.55% ± 0.99%), Sensitivity (96.10% ± 3.08%) and specificity (96.80% ± 1.48%).