A framework for histopathology image segmentation and classification
Metadata[+] Show full item record
[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] It is increasingly appreciated that tumor-stroma plays an integral part in cancer initiation, growth, and progression. Recently, it has been shown that the stromal elements of tumors hold prognostic as well as response-predictive information. Automated analysis of histopathology images have a great potential for both clinical applications (e.g., to reduce/eliminate inter- and intra-observer variations in diagnosis) and research applications (e.g., to understand the biological mechanisms of the disease process). This thesis proposed two computational image analysis frameworks to identify epithelial versus stromal tissue regions in images of Hematoxylin and Eosin (H and E) stained breast cancer specimens. The first framework used handcrafted image features with a support vector machine classifier. The H and E stained images were first segmented into coherent partitions/superpixels; then a number of regional color and texture features were extracted from these partitions; and finally a support vector machine classifier was trained to classify the image regions into epithelium and stroma classes. The second framework introduced a novel feature extraction and machine learning approach for identification of epithelium and stroma regions using deep learning methods. Deep convolutional neural networks (CNNs) were trained to extract hierarchical features from raw pixels of H and E stain images and to perform epithelium versus stroma classification. The classification results from deep convolutional neural networks were further fused with image segmentation results for improved performance. The proposed frameworks were evaluated on images from Stanford Tissue Microarray database. Deep learning based framework produced comparable results to the classification framework that used carefully hand-crafted image features. Experimental results showed that incorporation of an explicit segmentation step increased the classification accuracy in both frameworks.
Access is limited to the campuses of the University of Missouri.