Research On The Computer-Aided Diagnosis System Of Tumor Cells

This thesis is intended to develop a computer-aided diagnosis system of tumor cells based on Automated Vision Inspection techniques. Computer-aided diagnosis (CAD) of cell smears is realized by utilizing a color CCD video camera to capture cell images, making use of digital image processing and pattern recognizing techniques to classify cells. With a view to preprocessing of cell images, segmentation method of color images, classification and recognition of cells, a great deal of study is devoted and satisfying results are achieved.In the aspect of cell image preprocessing, algorithms in frequency field and linear and non-linear algorithms in space field are analyzed and compared with each other. Ultimately quick median filter algorithm is adopted to remove noise in the cell images.Color image segmentation is of great importance to the classification of cells. The thesis emphasizes on the research of segmentation algorithm of color images, which includes the segmentation of karyon and cytoplasm. Color spatial distribution of karyon, cytoplasm and background is statisticed and compared in the thesis. Referring to the relative theories of pattern recognition, the minimum distance-clustering algorithm is betaken to segment karyon in the color images. Aiming at characteristics of cytoplasm, Ostu thresholding algorithm is adored in I1I2I3 color space, and then adhesive cytoplasm is spitted by the way of watershed algorithm based on mathematical morphology.As for contour tracking, a contour tracking method based on characteristic code and retrospecting mechanism is present, which can track cell edge with short branch, so as to enhance its robustness. At last, the classification method is discussed in accidence and a classification method based on decision-tree is presented. In the method, the characteristic parameters of cells are extracted through different methods such as the morphologic, chroma and texture of cells, and then normal cells are eliminated gradually by the means which screens out normal cells from roughly to accurately, and in the end suspicious tumor cells are sorted according to their dubiety. The method can help pathologists work.By the CAD system presented in the thesis, 95.7% karyon and 93.8% cytoplasm are segmented successfully from cell images. The sorting result of suspicious tumor cells is correct in principle.