V-CoHOG: Volumetric Co-occurrence Histograms of Oriented Gradients for Texture Classification in Medical Imaging

Abstract

Image textures, which are properties that can describe the pixel intensities of an image, have been analyzed in order to perform a variety of tasks, including segmentation and classification. Features derived from textures which are invariant to changes such as contrast and transformations are of particular interest, due to their comparability across different images with related content. Hence, even early methods such as Gray-Level Co-occurrence Matrices (GLCM) have sought to describe textures using invariant features which proved to be effective in classification experiments. With the development of gradient based methods such as Histograms of Oriented Gradients (HOG) and derivatives including Co-occurrence Histograms of Oriented Gradients (CoHOG), an increasing variety of texture analysis methods have been applied to various domains, including medical imaging. When tasked with the classification of medical patients and healthy controls, methods such as Modified Co-occurrence Histograms of Oriented Gradients (M-CoHOG), which was applied to datasets of the neurodegenerative disorder Amyotrophic Lateral Sclerosis (ALS), have attained leading performance metrics. However, in the case of ALS, which can be difficult to diagnose, the lack of known biological imaging markers which could streamline the selection process of a region of interest (ROI), necessitates the contribution of an expert to segment the region of interest for M-CoHOG. Additionally, the consideration of only 2D imaging slices results in features with decreased spatial descriptiveness compared to volumetric approaches. In this thesis, a volumetric feature extraction method called V-CoHOG, extended on M-CoHOG, with automated ROI extraction is proposed. To reduce the feature vector size resulting from V-CoHOG, a feature selection method based on ReliefF is applied before classification using an ensemble model classifier with 5 base classifiers. The Canadian ALS Neuroimaging Consortium 1 (CALSNIC-1) and, for the first time, CALSNIC-2 ALS datasets are used for evaluation and comparison with four 3-dimensional convolutional neural network (CNN) methods, with the additional Alzheimer’s Disease Neuroimaging Initiative (ADNI) Alzheimer’s disease dataset used to demonstrate the versatility of the proposed method. Analysis of the results demonstrates that the proposed method is able to consistently outperform the 3D CNN approaches on ALS datasets, while approaching and sometimes outperforming M-CoHOG performance on the CALSNIC-1 datasets. Furthermore, a CUDA-accelerated implementation of M-CoHOG with significantly improved runtime performance on higher-resolution images is proposed. Finally, segmentation maps of selected features for ALS classification are overlaid onto the original imaging volumes. These maps allowed for the investigation of potential ALS imaging biomarkers, specifically their localization to brain regions present in the ROI.