Texture Analysis Using Wavelet-Based Multiresolution Autoregressive Model: Application to Brain Cancer Histopathology

Automated cancer diagnosis using histopathology images has evolved rapidly in the past few decades. This study focuses on tumor spatial heterogeneity in histopathology images. Long-range spatial dependencies in heterogeneous spatial process make the cancer diagnosis difficult and unreliable. A multiresolution autoregressive statistical model for histopathology images pertaining to brain cancer has been proposed. The primary idea is to study the complex random field and non-linear spatial interactions in a wavelet domain. Autoregressive parameters of vertical, horizontal and diagonal sub-band images represent a feature set for an image. SVM, MLP and fusion classifiers have been used to classify malignant samples. Classification accuracies of simple AR model and wavelet AR model against different model orders were compared. Better accuracies were obtained at lower model orders in wavelet AR. Introduction of wavelet transform to heterogeneous brain cancer images is a novel concept in model-based analysis and provides a new basis for analyzing histopathology images in computer-assisted diagnosis.