Performance Evaluation of Cancer Diagnostics Using Autoregressive Features with SVM Classifier: Applications to Brain Cancer Histopathology

Abstract

Until the recent past, cancer diagnosis was made using histopathology methods, where the pathologists study biopsy samples and make inferences. These inferences are based on cell morphology and tissue distribution which represent randomness in growth and/or in placement. These methods are highly subjective/arbitrary and can sometimes lead to incorrect diagnosis. Nowadays, computer-assisted diagnostic (CAD), based on very large database, can aid in objective judgment. This study emphasizes the contribution of a twodimensional (2D) autoregressive (AR) model for analysis and classification of histopathological images. In AR model, the parameters consist of a feature set of histopathological images obtained from biopsy samples taken from patients. These features are further used for analysis, synthesis and classification of cancer cells. The Yule-Walker Least Square (LS) method has been used for parameter estimation. The test statistics for the choice of a model order has also been suggested in this paper. It has been inferred that for a given sample image, the neighborhood is unique and solely depends on the properties of samples under consideration. Based on the features of AR parameters, samples are classified into two – healthy tissue and malignant tissue. The feature data sets have been classified using the linear kernel Support Vector Machine (SVM) classifier. In this work, we focus on measuring the performance of cancer diagnostic tests in terms of their recall, specificity, precision and F score. We observe that the fourth-order AR model gives promising results in performance evaluation using SVM classifier.