Deep Radiomic Analysis of MRI Related to Alzheimer’s Disease

Abstract

Alzheimer’s disease (AD) is the most common form of dementia, causing progressive impairment of memory and cognitive functions. Radiomic features obtained from brain MRI have shown a great potential as non-invasive biomarkers for this disease; however, their usefulness has not yet been explored for individual brain regions. In this paper, we hypothesize that distinct regions are affected differently by AD and, thus, that shape or texture changes occurring in separate regions can be expressed by different radiomic features. Moreover, to improve the classification of AD and healthy control (HC) subjects, we propose novel features based on the entropy of the convolution neural network (CNN) feature maps. The proposed approach is evaluated comprehensively using the Open Access Series of Imaging Studies database. Our experiments assess the significance of 45 different radiomic features from individual subcortical regions, via the Wilcoxon test. We also use the random forest classifier to identify the subcortical regions that best differentiate AD patients from HC subjects. Our analysis identified the features derived from several subcortical regions that show significant differences between AD and HC (corrected $p ). Specifically, we found correlation and volume features from the hippocampus (AUC = 81.19% – 84.09%) and amygdala (AUC = 79.70% – 80.27%) regions to have the greatest discriminative power. Furthermore, the proposed entropy features derived from CNN layers yielded the highest classification AUC of 92.58%, compared to 84.45% for the combined radiomic features of all subcortical. These results suggest that the proposed CNN entropy features could be used as an effective biomarker for AD.