A Robust Imaging Texture Descriptor for Analysing and Classifying Medical Images
University of Canterbury
Time & Place
Mon, 28 Sep 2020 15:00:27 NZDT in E16 - Engineering Core
Image texture features have demonstrated to be effective when characterizing image structure and classifying images into target categories. Extracting important and distinctive texture features from medical images brings some new challenges, as medical images contain very subtle texture differences and higher evaluation accuracy is required in clinical applications. This study investigated a popularly used local binary patterns (LBP) descriptor and its variant, local quinary patterns (LQP), and extended it to a new texture descriptor, rotation invariant LQP (RILQP). Such development aims to capture more texture patterns and to apply them in medical image classification. Two experimental tasks based on three medical imaging datasets were set up: classifying breast density into different categories in mammograms and grouping emphysema patches into three subtypes in CT images. A series of machine learning methods were employed to select optimum features and train the classification models based on the proposed texture descriptor. Competitive results were obtained in both classification tasks.
Haipeng Li is in his third year of PhD study and supervised by Prof. Mukundan and Dr. Shelley Boyd. His research work focuses on medical image analysis, feature extraction, image enhancement and classification.