摘要
目的针对宫颈癌筛查中需要从荧光宫颈图像和白光宫颈图像中检测出荧光图像中的多生暗区问题,提出一种基于灰度比值和水平集算法的分割方法。资料与方法首先对白光和荧光图像进行配准;然后通过k均值聚类算法从白光图像中分割出宫颈区域,并获得对应位置的荧白光的比值图像;最后在比值图像的基础上利用水平集方法实现多生暗区的分割。对宫颈区域的分割结果进行统计,以Jaccard Index(JI)为标准再次对3种方法(分别为基于配准图像的水平集算法的分割、基于无配准图像的水平集算法、基于配准图像的模糊C均值模糊聚类分割方法)的分割结果进行统计。结果采用基于配准图像的水平集算法的分割得到的平均敏感度比不做配准的水平集算法平均敏感度提高了13.32%,平均特异度提高了14.36%;比常用的模糊C均值聚类算法平均敏感度高了15.31%,并且平均特异度提高了14.40%。当采用JI精度指标时,基于配准图像的水平集算法比其他两种方法的平均精度分别高10.80%和18.19%。结论基于配准图像的水平集算法在荧光宫颈图像多生暗区的检测中是一种更为理想的分割方法。
Purpose To introduce a segmentation method based on gray ratio and level set algorithm regarding the detection of multidark area in fluorescence images from fluorescent cervical images and white light cervical images in cervical cancer screening. Materials and Methods Firstly, the registration of white-light images and fluorescence images was performed. Then, k-means clustering algorithm was adopted to segment cervical region from the white light images, and ratio images of the corresponding white light were obtained. Finally, the segmentation of multi-dark area was materialized using level set algorithm on the basis of ratio images. The segmentation results of the cervical region were gathered for statistical purpose. And with Jaccard Index （JI） serving as the standard, the segmentation results of the three methods （registration-based level set algorithm segmentation, non-registration-based level set algorithm, and registration-based FCM fuzzy clustering segmentation method） was harvested the second time for statistics. Results The mean sensitivity obtained from registration-based level set algorithm segmentation was 13.32% higher than that of the non-registration level set algorithm, with mean specificity improved by 14.36 %, and was 15.31% higher than that of the FCM clustering algorithm, with mean specificity raised by 14.40%. When applying the （JI） precision index, the registration-based level set algorithm was 10.80% and 18.19% higher than the other two methods in terms of mean precision. Conclusion The registration-based level set algorithm is a more ideal segmentation method in multi-dark area detection of fluorescent cervical images.
作者
刘君
余婷婷
石慧娟
陆晗
LIU Jun;YU Tingting;SHI Huijuan;LU Han(Key Laboratory of diangxi Province for Image Processing and Pattern Recognition,Nanchang Hangkong University,Nanchang 330063,China)
出处
《中国医学影像学杂志》
CSCD
北大核心
2018年第9期695-702,共8页
Chinese Journal of Medical Imaging
基金
国家自然科学基金(61402218)
江西省自然科学基金(20151BAB205050)
江西省教育厅基金(GJJ14503)
关键词
宫颈肿瘤
阴道镜检查
荧光
算法
图像处理
Uterine cervical neoplasms
Colposcopy
Fluorescence
Algorithms
Image processing
