The Histological Image Analysis For Developing Computer-Aided Diagnosis On Breast Cancer

Nuclei detection and color unmixing are two important problems in histopathological images analysis. In this paper, a stacked sparse autoencoder (SSAE) based feature extraction and sliding window method was proposed for rapid, efficient and accurate detection of nuclei from high resolution pathological images. High-level feature was extracted from the training samples which include both nuclei and non-nuclei patches using SSAE. The supervised Softmax classifier was trained with these high-level features. The trained classifier was employed for automated nuclei detection in histopathological images. During the detection, sliding window was used to select patches from these high resolution pathological images. After extracting high-level features with SSAE, these patches are recognized by Softmax classifier to decide to be cells or not. If some patches are recognized as cells, corresponding positions will be marked as cells. In order to verify the effectiveness of the method on detecting nuclei on histopathological images, this dissertation compared the proposed method with other models based on cell segmentation, for example Expectation-Maximum Algorithm, Blue Ratio Thresholding and Color Deconvolution, for nuclei detection. The experiment results showed that the proposed method achieved highest Precision, Recall and F1-measure values, which were 71.5%,82.3% and 76.5% respectively. Also, this dissertation presented a new unsupervised Sparse Non-negative Matrix Factorization (SNMF) based approach for color unmixing on breast histopathological images. The unmixing images were handled with morphological processing for cell segmentation. The traditional methods for color unmixing were based on color deconvolution which needs specific knowledge of the data, however, this is not required in the proposed method. Compared with the current Non-negative Matrix Factorization(NMF) based methods and other methods, such as Principle Component Analysis (PCA), independent Component Analysis (ICA), Expectation-Maximum Algorithm(EM) and Blue Ratio Thresholding(BRT), SNMF with sparse constraint shows better segmentation accuracy.