Automated Image Analysis for Nuclear Morphometry Using H&E and Feulgen Stains in Prostate Biopsies
thesisposted on 2014-04-15, 00:00 authored by Kusuma Bapure
The thesis addresses the problem of analyzing prostate cancer biopsy image data. It describes an unsupervised semi-automated method for segmentation of Epithelial Nuclei in Hematoxylin and Eosin (H&E) stained Prostate Biopsy images and investigates its use in modeling the data and its effectiveness predicting cell alterations. Existing methods have largely focused on the use of Feulgen-stained prostate cancer biopsies in the analysis of Nuclear Morphometry due to its DNA staining capability. In this thesis the potential for the use of the more easily available H&E staining is investigated since H&E stains are widely used in medical diagnosis due to the simplicity in the staining procedure. Our results provide evidence that the H&E can yield a performance comparable to a reference method that uses Feulgen-stained biopsies. In this study, a set of over 180 features are extracted from each image in a database of 42 Hematoxylin and Eosin (H&E) stained negative prostate biopsies. Grouped as Cases (subjects with no cancer on their initial biopsy and subsequently received a cancer diagnosis) and Controls (followed for an equivalent period of time without cancer being detected). Prostate biopsies stained with H&E are studied as test and training samples in comparison with the Feulgen stained slide results. A robust algorithm, implemented in Matlab, has been developed for semi-automated segmentation of glandular structures in histopathology imagery using an Aperio ScanScope. K-Means clustering algorithm and other morphological operations are used to pre-process the images to filter out irrelevant structures. The nuclei centers obtained with Radial Symmetry Transform act as regional minima for Marker Controlled Watershed Segmentation. The approach detects multiple nuclei from a closely spaced/merged cluster of nuclei. Architectural and texture features were measured for each cell image. The method provides good performance in terms of segmentation accuracy. In this preliminary study, the good agreement between the morphometric results and general histomorphologic data demonstrates the importance of nuclear morphometric analysis using H&E stains in benign prostate biopsies, which could be extended to other cancer types. The person-level Multi-Feature Score (pMFS) is produced by applying Logistic Regression to the reduced feature set. Its result is verified by an Area-Under-receiver-operating-Curve (AUC) value of 0.77.