Multi-scale radiomic analysis of sub-cortical regions in MRI related to autism, gender and age
We propose using multi-scale image textures to investigate links between neuroanatomical regions and clinical variables in MRI. Texture features are derived at multiple scales of resolution based on the Laplacian-of-Gaussian (LoG) filter. Three quantifier functions (Average, Standard Deviation and Entropy) are used to summarize texture statistics within standard, automatically segmented neuroanatomical regions. Significance tests are performed to identify regional texture differences between
... vs. TDC and male vs. female groups, as well as correlations with age (corrected p < 0.05). The open-access brain imaging data exchange (ABIDE) brain MRI dataset is used to evaluate texture features derived from 31 brain regions from 1112 subjects including 573 typically developing control (TDC, 99 females, 474 males) and 539 Autism spectrum disorder (ASD, 65 female and 474 male) subjects. Statistically significant texture differences between ASD vs. TDC groups are identified asymmetrically in the right hippocampus, left choroid-plexus and corpus callosum (CC), and symmetrically in the cerebellar white matter. Sex-related texture differences in TDC subjects are found in primarily in the left amygdala, left cerebellar white matter, and brain stem. Correlations between age and texture in TDC subjects are found in the thalamus-proper, caudate and pallidum, most exhibiting bilateral symmetry. Autism is a complex developmental disability that often appears during infancy, typically in the first three years of life 1 . It is a spectrum disorder that affects about one in 300 children, with individuals affected differently and to varying degrees 2 . The causes of autism are not yet fully understood, and a combination of developmental, genetic, and environmental factors are believed to be involved 3,4 . With the development of in-vivo brain imaging technologies such as Magnetic Resonance Imaging (MRI), significant progress has been made toward understanding the physiological characteristics of Autism Spectrum Disorder (ASD). Morphological analysis methods have identified links between regional image measurements and ASD, shedding some light on the mechanisms of this complex disorder. Such methods typically quantify physiological properties of neuroanatomical structures from image data, e.g. volume, thickness, shape, etc., and then identify regions/features for which these properties exhibit statistically significant differences between subject groups (e.g., normal control and ASD) or correlations with variables of interest (e.g., age). A number of studies have identified physiological differences between ASD and healthy subjects in several key brain regions, including the putamen 5 , cerebellum 6,7 , hippocampus 8,9 , amygdala 10 , and corpus callosum 11 . The majority of these studies have focused on regional volume or intensity measurements, however, and have not fully exploited the rich information contained in MRI. A recent body of research, commonly referred to as radiomic analysis, focuses on texture and shape features derived from image data, that provide a source of information complementary to traditional voxel-wise or volumetric measurements. In general, radiomic analysis hypothesizes that texture operators tuned to an appropriate spatial extent or scale may provide information regarding the microstructure of the biological tissues observed, a hypothesis closely tied to results from scale-space theory 12 . Texture features can be computed by various means including linear filtering operations (e.g., Laplacian-of-Gaussian (LoG) 13 , wavelets) or gray-level co-occurrence matrices (GLCM) 14 , and offer a means of characterizing localized image variations arising from tissue heterogeneity, boundary smoothness 15 , etc. Such features have been successfully applied in a variety of image analysis contexts, including segmentation and computer-aided diagnosis        . In lung and head-and-neck cancer data, radiomic features were shown to have significant prognostic power, their signature relating intra-tumor heterogeneity with gene-expression patterns 23 .