Study On The Recognition Of Breast Tumor Tissue Microarray Based On Microscopic Hyperspectral Imaging

In recent years,the incidence of breast cancer has been increasing,which is an enormous threat to women’s physical and mental health.At present,histopathological examination is the golden standard for the diagnosis of breast cancer,which is mainly performed by doctors with microscopic examination of pathological section.This process is time-consuming and can be influenced by doctors’ subjective factors and experience.Traditional digital image based computer-aided diagnosis can provide certain recognition and analytical results for pathologists with finite information,limiting the improvement of accuracy.Microscopic hyperspectral imaging technology has the image-spectrum merging characteristics,which is a new way for the recognition of breast tumor.It can provide not only morphological and structural information of tissues and cells but also spectral information.The morphological features of breast cancer nests and cells are important references for evaluating the degree of tumor deterioration.In this thesis,microscopic hyperspectral imaging technology is applied in the identification and quantitative analysis of breast tumor tissue microarray and the segmentation algorithms of cancer nests and cells are studied.Firstly,in order to take full advantage of the spatial and spectral features of hyperspectral images for cancer nest segmentation,support vector machine based on combination of spatial-spectral features(SVM-CSS)is proposed to identify cancer nest via pixel-by-pixel classification.To further improve the recognition efficiency,a U-net segmentation model based on principal component analysis(PCA-Unet)is established.In this method,PCA is employed to extract spectral features of microscopic hyperspectral images.The U-net deep learning framework is adopted to achieve endto-end image segmentation,which lays the foundation for quantitative description of relevant morphological features.Secondly,watershed algorithm based on combination of spatial-spectral features(WCSS)is proposed for cell segmentation.In this method,the target abundance distribution is extracted via endmember extraction and spectral unmixing.Then the marker-based watershed transform is applied to complete segmentation.Finally,the cancer nest cells and normal cells can be divided by further binding to cell and cancer nest segmentation results.Then the parameters of tumor tissue morphology are measured to provide quantitative reference indices for doctors’ pathological diagnosis.The experimental results show that PCA-Unet has more precise segmentation in cancer nest than SVM-CSS and other microscopic hyperspectral image recognition methods,reaching 87.14% accuracy of pixel recognition on the acquired dataset.In the field of cell segmentation,it gains better results by using WCSS based on endmember extraction of automatic target generation process and spectral unmixing of fully constrained least squares,which reaches 90.68% accuracy of pixel recognition on our dataset.Automatic recognition of breast tumor can be achieved by using the identification algorithms of breast tumor tissue microarray based on microscopic hyperspectral imaging,providing a new method for pathological diagnosis of breast cancer.