A Computer-aided Diagnosis System for Automated Gleason Grading of Prostatic Adenocarcinoma from Digitized Histology
Scott Doyle BS; Rutgers University; John Tomaszewski MD; University of Pennsylvania; Michael Feldman PhD; University of Pennsylvania; Anant Madabhushi PhD; Rutgers University;
Content:
We present a computer-aided diagnosis system to automatically assign a Gleason grade to digitized prostate tissue images. We quantify the size, shape, and location of the glands and nuclei by extracting a set of features from each image. A classification algorithm uses these extracted features to determine the Gleason grade in the image. This system will improve rates of inter- and intra-observer variability in grading by creating quantitative signatures of tissue phenotype. In this study, we focus on distinguishing between Gleason grades 3 and 4, since intermediate grades are a clinically relevant problem and are the greatest source of clinical variability.
Technology:
A set of Hematoxylin and Eosin stained prostate tissue slides are scanned into a computer using a whole-slide digital scanner. These images are analyzed using a set of routines implemented in the MATLAB (MathWorks, Natick, Massachusetts, United States) software package.
Design:
We use image analysis methods to extract the centroids of the nuclei and the glands within a digital image of prostate tissue, as well as the boundary of each gland. A number of features are computed from the centroids and boundaries that can quantify the architecture (spatial location) and morphology (size and shape) of the tissue structures. In addition, we calculate texture features such as Haralick and wavelet filter features to further quantify the tissue region. We then use manifold learning techniques such as graph embedding to visualize these image regions in a low-dimensional reduced feature space. Finally, we apply a classification algorithm to classify the tissue image.
Results:
The results of applying the manifold learning technique are shown below in Figure 1. We illustrate each image as a point on the feature space reduced by graph embedding. Points appearing near each other in the reduced space have similar feature values, and likely belong to the same class. These plots illustrate the ability of our feature set to distinguish between these image regions according to Gleason grade. When a classification algorithm is applied to this data, we obtain classification accuracies of 95.8% when distinguishing between Gleason grades 3 and 4, 96.2% between grade 3 and benign epithelium, and 100% between grade 4 and benign epithelium.
Conclusion:
We have presented a CAD system for the automated grading of prostate cancer from digitized histology. Our results indicate that our set of morphological, architectural, and textural image features can accurately distinguish between Gleason grades of tissue with a high degree of accuracy.
