Automatic Segmentation And Recognition Of Breast Histopathology Images Based On Wavelet Decomposition And Multi-scale Morphology

Breast cancer has become the most common cancer in women. There has also been an increase in the incidence of breast cancer year by year. Early discovery combining with early diagnosis is the key to improve curative effect of breast cancer. The only way to make a definite diagnosis of breast cancer is biopsy pathologic examination. Observing breast slice images under a microscope we can find that most cancerous cells are darker and bigger and have higher brightness variation than normal cells. At present, the effective and practical way to make quantitative analysis of microscopic cell images is using computer technology, which can assist doctors to discriminant cell pathological ill state. Due to some inevitable affection like the staining and illumination in image collection process, the breast pathologic images have such problems as uneven dyeing, inconspicuous contrast between cells and background, existing holes inside cells, overlapped cells, etc.For the multiple existing problems in the human breast cell images, this thesis proposed an automatic segmentation algorithm based on wavelet decomposition and multi-scale region-growing(WDMR) combining with double strategy splitting model(DSSM) and put forward a feature selection algorithm based on hypothesis testing and the corresponding automatic classification. The main contents are as follows:① By observing the characteristics of human breast cell image, this thesis studied and proposed a pre-segmentation method combining wavelet decomposition with multi-scale region-growing(WDMR). The seed points of region growing were the erosion results of wavelet decomposition by different levels and the criteria was the different threshold value of grayscale. Multi-scale region growing results were obtained by different seed points with different criteria. Followed with the optimal voting mechanism to choose best cell regions.② In order to ensure the accuracy of the segmenting line and the efficiency of the algorithm, a double strategy splitting model(DSSM) is applied to split overlapped cells. This model separated different degree of adhesion cells by different methods. The adaptive mathematical morphology was utilized to separate less overlapped cells and Curvature Scale Space(CSS) corner detection method to separate heavier overlapped cells.③ To realize classification and recognition of cancer and normal samples, the thesis studied and proposed a morphological and texture features extraction method for every segmented cell, then selected features which have significant difference by statistical hypothesis test method. Finally verified the effectiveness of the proposed feature extraction and selection algorithm.The automatic segmentation and feature extraction algorithm for breast cell images proposed in this paper provided a new train of thought, and provided a new theoretical and method basis for auxiliary diagnosis of breast cancer based on breast histopathology images, and contributed to the clinical application of breast histopathology images analysis. The proposed method has been proved to have certain theoretical significance and application value.