| Intelligent cervical cell image analysis is one of the specific tasks that applies AI technology to auxiliary modern medical diagnosis.Recently,many researchers are dedicated to the research on intelligent cervical cell image analysis and have made remarkable progress.However,due to the characteristics of realistic cervical cell images as well as cervical cell's complex shapes and structures,the accuracy and real-time performance of existing cervical cell image processing,detection and recognition technologies need to be improved,in order to meet the public's pressing demands for intelligent cervical cell image analysis.By the virtue of pattern recognition and machine learning theory,this dissertation provides a thorough analysis to image segmentation,feature extraction,feature selection,feature combination and classification of cervical cell image,and conducts in-depth research on the key technologies of intelligent cervical cell image analysis to promote the accuracy and speed of the intelligent analysis system for cervical cell images.This dissertation makes the following contributions:(1)To study on the segmentation of non-overlapping single cervical cell color image,we have proposed an accurate and fast image segmentation method based on super pixel Gap-search MRF(Markov Random Field)model.Based on the pre-segmented super pixels,this model can extend MRF model to label super pixel patches to partition the image into cell nucleus,cytoplasm and background directly.Moreover,due to the proposed Gap-search mechanism,a large amount of redundant computation is reduced to solve the problem and the proposed method is much faster than pixel-based MRF model.Experimental results show that the proposed segmentation method based on super pixel Gap-search MRF can segment the non-overlapping cervical single cell images more accurately and quickly than other segmentation approaches.(2)The segmentation of cervical cell overlapping images is explored and this problem is divided into two categories: “partial overlapping” segmentation problem and “true overlapping” segmentation problem.1? For “partial overlapping” segmentation of cervical cell image,we propose an improved method based on super pixel and K-means++.Firstly,SLIC super pixel algorithm technique is employed to divide the filtered image into super pixel patches.Secondly,we extract 13 dimension features from the divided super pixels and these super pixels are clustered into 4 classes by K-means++ algorithm.According to the different intensities of cluster's centres,the image is segmented to cervical nucleus,background,candidate cytoplasm and overlapped region.Finally,in the light of the prior knowledge of the cell shape,candidate cytoplasm and overlapped regions,cytoplasm is identified by removing the outside of cell shape.Experimental results show that the proposed method performs better than other classic unsupervised methods for the segmentation of “partial overlapping” cervical color image segmentation.2? For“true overlapping” segmentation of cervical cell image,we propose an improved method which combines graph cuts and Voronoi diagram algorithms.Firstly,this method employs graph cuts for scene segmentation to identify background and foreground.Secondly,after detecting nuclei regions of each cell clump as seeds of Voronoi diagram,non-overlapping“rough” cells are segmented from the clump.Thirdly,according to the cell prior shape,the overlapping “compensation” region of each rough cell,which is not in Voronoi cell,is extracted.Finally,we combine “rough” cell region and “compensation” cell region as a whole cell.Experimental results show that the proposed method achieves the same segmentation accuracy as existing best segmentation methods over two challenge datasets of overlapping cervical cell segmentation.(3)The problem of feature selection and combination for cervical cell images is studied.We have proposed margin based adaptive feature selection and combination method.The basic idea is proposed based on minimizing the distance among instances in a cluster,meanwhile maximizing the distance among instances belong to different clusters.Specifically,the prior probability about the training data is used when solving multivalue weighted vector for feature optimization.In addition,when the feature selection and combination are conducted,the original features are transformed into the new feature space.It thus becomes easier to conduct classification of the instances in the new feature space.Experimental results illustrate that the proposed feature selection and combination method outperforms its counterparts on accuracy when performing the abnormal detection for cervical cell images.(4)This dissertation studies the outlier detection for cervical cell images.It argues that cervical cancer cell detection system(C3DS)should guarantee both the accuracy and the efficiency.1? To improve the accuracy of a C3 DS,this dissertation proposes a Threephase outlier detection model for cervical cell image analysis.The model separately optimizes the three stages in cervical cancer outlier detection based on a well-segmented cell image.Specifically,it first extracts 160 features to represent the intrinsic characteristics of the cervical cell.Then,a margin-based adaptive feature selection method is used to learn an optimal combination of these features for outlier detection task.After that,the optimized features are fed into a Two-stage classified strategy for cancer cell detection.Experimental results demonstrate that the proposed model can achieve 96.98%detection accuracy,which outperforms other 16 compared models.2? Regarding to promote C3 DS accuracy and efficiency,this dissertation proposes an outlier detection model based on MI-ELM.The proposed model uses multi-instance learning to integrate feature selection and combination with multi-class classification,and adopts MI-ELM to classify the normal and abnormal cell directly.Here,cervical cell features are divided into three categories,i.e.color,shape and texture,according to their semantical meanings.Then,to enable MI-ELM,it treats a cervical cell as a “bag” and a category of the features as an “instance”.Finally,the well-trained MI-ELM is employed to identify abnormal cells from normal cells.Experimental results show the proposed method can not only speed up the cervical cell outlier detection but also improve the accuracy compared with other competitors.(5)To classify the cervical cells into four categories,i.e.normal squamous cell,normal columnar cell,abnormal cervical cell and cervical cancer cell efficiently and accurately,we propose a PCA(Principal Component Analysis)-based two-cascaded method.Specifically,in our method,3 SVMs are combined under the guidance of OVO strategy as base classifiers to classify cervical cells into three categories in the first stage.In the second stage,1 SVM classifier is applied onto the abnormal sub-group to differentiate more similar abnormal cervical cells and cervical cancer cells.Besides,PCA is conducted before each classification procedure to reduce the dimensionality and filter the noisy information in features to improve the performance on speed and accuracy of the method.Experimental result demonstrates the superiority of our proposed method on both classification accuracy and time consumption against previous approaches. |
PDF Full Text Request