Research On Early Tumor Screening Method Based On Computer Vision

Automated processing and quantitative analysis of the medical image has become an important area in image processing and biomedical engineering fields, due to the rapid advancement of medical imaging technology. As an important subfield of medical image processing, quantitative analysis, processing and recognition of microscopic cell images has been put forward by researchers long ago and has become hot issue recently.Aiming at the cell images obtained from Feulgen stained cytological samples, this dissertation focuses on the quantitative cytological analysis and automated early cancer diagnosis technology. By combining cytological pathology knowledge, image processing technology and computer vision theories, we conduct a systematic and comprehensive study on the technologies of cell nuclear segmentation, quantitative cytological analysis and feature extraction, and automated cell recognition and classification. The main contributions of this thesis are given below.(1) An active contour based supervised cell nuclear segmentation algorithm is proposed. Based on the comprehensive study and comparison of the threshold method, watershed method and cell nuclear boundary relocation algorithm, we combine unsupervised active contour method, directional gradient and support vector regression algorithm to improve the accuracy of the cell nuclear segmentation. The experiments show that the proposed method has encouraging segmentation results and stable performance.(2) Considering different kinds of cell nuclei can be identified by its shape, border, texture and optical characteristics, systematic study of the feature extraction methods is conducted. To effectively solve the problems caused by the uncertain and uneven nature of the staining and background light intensity, we proposed a normalization method to calculate the optical density image and normalized light intensity image, which can be used to replace the original grayscale image in the feature extraction process. By using the proposed method, we improve the stability and robustness of the extracted features.(3) Due to some unavoidable factors, certain errors may occur in the measurement of the DNA content through processing and analysis of the cell nuclei image. A new calibration approach based on morphology and SVR regression algorithm is proposed. Through spatial and optical calibration, the approach sufficiently utilizes the information hidden in the images and features of the cell nucleus. The experiment shows that the approach could increase the measurement accuracy significantly, and show positive significance in improving the specificity and sensitivity of the pathological diagnosis.(4) A method to reduce the dimensionality of cell nuclei feature vectors is proposed. First, the statistically based F-score value of each cell nucleus feature is calculated and apparently useless features are rejected. Then, the random forest algorithm was conducted on the remaining features, and they were sorted in descending order by RF-score value. After evaluating the performances of the cell nuclei classifiers under the conditions of different numbers of features, the final feature vector for cell nuclei recognition was determined. Experimental results show that compared with the original cell nuclei classifier, the dimensionality reduction algorithm can save the computation time and raise the cell nuclei recognition accuracy significantly.At the end of this thesis, the performance of the diagnosis technology based on the methods mentioned above is evaluated.Finally, we summarize the research work for the dissertation, and discuss some research topics and directions related to this work in the future.