Research On Algorithms Of Peripheral White Blood Cell Segmentation And Recognition

This dissertation studies the algorithms of White Blood Cell (WBC) segmentation and recognition. WBC plays a vital role in human immune system. In clinical practice, haematologist diagnose blood system diseases by observing the number of different type of WBC and the change of WBC’s morphology. Hence, the count of WBCs is one of significant in-dexs in blood routine examination. WBC morphology system is designed to automate the traditional microscopy by using segmentation and recogni-tion algorithm in image processing and pattern recognition. Due to WBC’s diversified types and morphology, WBC segmentation and recognition is a difficult problem.This dissertation studies WBC segmentation using multi-feature non-linear combination, WBC segmentation based on WBC location, and WBC recognition based on deep convolution neural network. The main contents are as follows:1. Improve the algorithm of WBC segmentation based on multi-feature nonlinear combination. As WBC have definite region of interest, multi-feature nonlinear combination is more suitable to WBC segmentation. In order to simplify the training process of the parameters in nonlinear com-bination model, we construct an intermediate target between training tar-get and the output of nonlinear combination model, and the parameters are trained by an iteration process. The experimental results show that non-linear iterative learning is an effective method. In the process of evaluat-ing WBC segmentation algorithm, we define a boundary-based F-measure. Compare with region-based F-measure, boundary-based F-measure is more effective for images that have definite region of interest.2. We propose the segmentation algorithm based on WBC location, for the interference of irrelevant regions (red blood cell and platelet) in the process of segmentation. That’s to say, we first detect the location of WBC, and then segment WBC in subimage which contain a complete WBC. In regard to WBC location, the detection windows are fixed in several certain sizes based on the sizes of WBCs. Then, we locate the position of WBC accord to the feature of detection windows. For nuclear segmentation, we propose an adaptive thresholding segmentation algorithm. We estimate the threshold depended on the intensity histogram of subimages which contain a complete WBC. Experimental results show that adaptive thresholding is not only an effect segmentation method, but also robust to different WBC datasets. As for cytoplasm segmentation, we use the location of WBC, which is the result of our detection algorithm, as a background label which is need manual labelling in the GrabCut algorithm to reduce user interaction and to implement automatic cytoplasm segmentation.3. Most existing WBC classification algorithms extract features based on haematologist’s recognition experiences, and then use some classifiers to classify the WBCs. However, due to the complex nature of WBCs and the difference between feature descriptors and haematologist’s recognition experiences, the existing WBC classification algorithms are not suitable for WBCs with more categories. Therefore, we need a more powerful tool to represent and classify WBCs. We introduce a deep convolution neural network into WBC recognition, and design a deep convolution neural net-work. Experimental results show that our deep convolution neural network is not only superior to the existing WBC classification algorithms, but also have room for improvement.