Abstract for HONS 03/21
Texture-Based Segmentation for Applications in Computational Pathology
Wikke Nijhof
Department of Computer Science and Software Engineering
University of Canterbury
Abstract
Nuclei segmentation is an important step in the diagnosis of cancer, since tissue and nuclei properties are different in each disease stage. Recent developments in the field of computational pathology have led to the rapid increase in computerised methods for nuclei segmentation and classification. However, one weakness in current segmentation models is that they use intensity, colour and shape as primary features with comparatively less importance given to texture information. We perform a detailed investigation of common texture descriptors and a statistical analysis of Local Binary Patterns (LBP) and Grey Level Co-occurrence Matrices (GLCM), allowing us to propose the best texture features for nuclei segmentation. We show that colour features alone do not provide enough information for accurate nuclei segmentation and that the inclusion of texture descriptors is useful in improving classification accuracy. Our best method combines texture features gathered using GLCM combined with colour features and achieves an overall Dice Coefficient of 0.786 on the Mo-NuSeg dataset [1].