Higher Order Statistical Analysis in Multiresolution Domain - Application to Breast Cancer Histopathology

Objective is to analyze textures in breast histopathology images for cancer diagnosis. Background: It is observed that breast cancer has second highest mortality rate in women. Detection of cancer in early stages can give more treatment options and thus reduce the mortality rate. In cancer diagnosis using histopathology images, histologists examine biopsy samples based on cell morphology, tissue distribution, randomness in their growth or placements. These methods are time taking and sometime leads to incorrect diagnosis. These methods are highly subjective/arbitrary. The new techniques use computers, archived data and standard algorithms to provide fast and accurate results. Material & Methods: In this work we have proposed a multiresolution statistical model in wavelet domain. The primary idea is to study complex random field of histopathology images which contain long-range and nonlinear spatial interactions in wavelet domain. This model emphasizes the contribution of Gray level Run LengthMatrix (GLRLM) and related higher order statistical features in wavelet subbands. The image samples are taken from `BreaKhis' database. The standard database generated in collaboration with the P&D Laboratory-Pathological Anatomy and Cytopathology, Parana, Brazil. This study has been designed for breast cancer histopathology images of ductal carcinoma. GLRLM feature dataset further classified by SVM classifier with linear kernel. The classification accuracies of signal resolution and multiresolution have been compared. Results: The results show that the GLRLM based features provides exceptional distinguishing features for multiresolution analysis of histopathology images. Apart from recent deep learning method this study proposes use of higher order statistics to gain stronger image features. These features carry inherent discriminative properties. This higher order statistical model will be suitable for cancer detection. Conclusion: This work proposes automated diagnosis. Tumor spatial heterogeneity is the main concern in analyzing, diagnosing and grading cancer. This model focuses on Long range spatial dependencies in heterogeneous spatial process and offers solutions for accurate classification in two class problems. The work describes an innovative way of using GLRLM based textural features to extract underlying information in breast cancer images.