Computerized Discrimination of Breast Cancer Grade from Digitized Histology via use of Architectural and Textural Features
Electronic Poster - Award to Submit Full Manuscript on Behalf of APIII to Cancer InformaticsBhakti Tulpule MSEd; Rutgers University; John Tomaszeweski MD, PhD; University of Pennsylvania; Michael Feldman MD, PhD; University of Pennsylvania; Anant Madabhushi PhD; Rutgers University;
Content:
We present an image analysis methodology to distinguish between high and low grades of breast cancer from digitized breast histology. Image based features representing texture and architecture of the tissue and distribution of nuclei are extracted from histology images. Manifold learning followed by classification is used to distinguish between high and low cancer grades. The goal of this work is to reduce inter- and intra-observer variability among pathologists in histological grading of breast cancer.
Technology:
A set of Hematoxylin and Eosin stained breast tissue slides are scanned into a computer using a whole-slide digital scanner. A total of 20 images (10 belonging to high grade, 10 to low grade) have been used. These images are analyzed using a set of routines implemented in the MATLAB software package.
Design:
From the digitized histology we extract 26 architectural features through constructing three graphs, Voronoi, Delaunay, and the Minimum Spanning Tree, by connecting identified gland and nuclear centers within the tissue. These features capture the architecture of the tissue. We extract 25 nuclear features which capture the distribution of nuclei. A total of 13 Haralick features were also utilized to quantify the textural variations in cancer grades. We then use manifold learning, a non-linear data dimensionality reduction scheme, called Isomap to distinguish between images corresponding to low and high grades in a low-dimensional embedding space. A Support Vector Machine (SVM) classifier is then applied to the Isomap output to group the tissue images into two classes.
Results:
The result of applying the manifold learning technique is shown in Figure 1. The feature set is reduced from 64 to 2 dimensions. In this reduced dimensionality space each image is represented as a point. Points closer to each other are expected to belong to the same grade. A good separation is observed between the two sets of images and hence between the two grades. SVM classifier applied to the Isomap output gave a classification accuracy of 100% in distinguishing between high and low grade breast cancers.
Conclusion:
Using architectural, textural and nuclear features in conjunction with manifold learning has enabled us to distinguish between high and low grades of cancer with 100% accuracy using the SVM classifier.
