| The American Cancer Society estimates that in 2012 about 577,190 people will die of cancer in the United States (US). It is estimated that in 2012, 1,638, 910 people will be diagnosed with cancer, which remains the second most common cause of death in the US. NIH estimates from 2007 put the overall cost of cancer that year as ;This dissertation contains a step wise progression of in situ -in vivo approaches to various challenges in pathology applications towards a proposed combination of multi-spectral imaging methods and image analysis techniques to create a prototype automated computer-aided system towards the diagnosis of cancer using digitized multispectral slides. The techniques have been applied to many areas from fresh stained and unstained breast tissue to in vivo imaging of lesions suspected to be melanoma. There are several original contributions: first, a quantitative assessment of the utility of various multispectral devices and imagery for segmentation and classification tasks in pathology. Next, tissue level and object level segmentation algorithms are developed for various histological classes along with quantitative metrics. In addition, references of both tissue, spatial, and object level features are extracted to create a comprehensive feature selection framework for classification of objects and images. The tools, algorithms, and methods described are for quantifying molecular changes in light microscope images of cellular structures indicative of cancer or precancerous lesions.;For the cervical and melanoma applications object level features as implemented are versatile and useful to extract important features even from relatively inaccurately segmented images. In addition, the use of non-nuclear features, like features of the cytoplasm and stroma has very good classification performance when compared to commercial devices. The system is in two parts: the segmentation of squamous epithelium and the subsequent diagnosis of CIN. For the segmentation of squamous epithelium, to save processing time, a multi-resolution method is developed to segment cervical virtual slides.;The nuclei segmentation method uses robust texture features in combination with a Support Vector Machine (SVM) to perform classification. Medical histology rules are finally applied to remove misclassifications. In tests using 31 virtual slides, the segmentation achieves an average accuracy of more than 94.25%. Training nuclei are spectrally classified into Normal, CIN I, CIN II and CIN III. The final diagnosis for a slide region is based on combining the classification of nuclei and classical morphologic features. .The robustness of the system in terms of regional diagnosis is measured against slides manually classified by three pathologists. Results indicate that the multispectral imaging system offers a promising basis for a computer-assisted diagnostic tool. Its main limitation is seen to be in the selection of more extensive and more varied training data. |
