White Matter Lesions and Pattern Recognition in MRI of Neurodegenerative Dementia

Abstract

Introduction Expected age is increasing globally and dementia is a common outcome for an increasing number of people. Dementia is a demanding syndrome for the patient and the environment as well as it is costly for society. Damaging changes to the cerebral blood flow also called white matter lesions (WML) are common in the elderly and is expected to increase as age advances. It has been reported that these types of lesions affect cognition in healthy elderly. They are also associated to Alzheimer’s disease but have not been much studied in DLB. Quantitative analysis and machine learning have a potential to contribute in understanding the disease process as well as aid in diagnosis. Methods Quantitative analysis of WML volumes were calculated using an automatic segmentation routine on magnetic resonance images (MRI) of subjects with Alzheimer’s disease (AD), Lewy body dementia (LBD), and normal controls (NC). Statistical tests were performed to compare groups as well as to investi- gate relations to cognition. Additionally, WML volumes were used as features in a machine learning (ML) environment to check whether WML volume were able to classify subjects with AD and LBD from NC. Texture analysis (TA) may be able to document changes at a microstructural level and was performed in WML an non-WML regions of the different types of MRI’s (FLAIR and T1). 2D- and 3D TA features were calculated and used in classification with the aim to serve as a tool for computer aided diagnosis (CAD) in dementia. The dataset used was imbalanced meaning that the number of subjects in each group were very different. Two methods for handling the imbalanced data were tested, namely upsampling and cost-sensitive classification.

Results and conclusions Severity of WML did neither differ significantly between subjects with dementia and NC nor between mildly demented patients with AD and LBD. WML severity were associated with cognitive decline in AD, but not LBD suggesting that WML contributes to cognitive decline in AD, but not LBD. More studies of the potential clinical impact of WML in patients with LBD are needed. The best classification results obtained using WML volumes as features in an ML framework discerning subjects with dementia from healthy controls were an area under curve (AUC) of 0.73 and 95% confidence interval of 0.57 to 0.83. We experienced better classification results when using TA features compared to WML volumes in classification and better results when performing classifi- cation on TA features calculated from T1 MRI compared to FLAIR MRI. A total accuracy, reported as mean with standard deviation in brackets over cross validation folds, of 0.97(0.07) or higher was reported for the dementia vs. NC, AD vs. NC, and LBD vs. NC classification problems for both the 2D- and 3D texture analysis approaches. In the AD vs. LBD case a total accuracy of 0.73(0.16) was reported using the 2D TA approach slightly exceeded by the 3D TA approach were 0.79(0.15) was reported. It seems like the results do not differ much when performing analysis in different regions of the brain and that the results vary in an inconsistent way. Using upsampling increased classification accuracy to a large extent in the LBD class at the expense of total accuracy and the accuracy of the AD class. In both the two-class problems NC vs. AD and NC vs. LBD, adding cost- sensitivity increased classification performance in many of the tests, but upsam- pling increased accuracy even more in most of the tests. High classification performance was achieved when classifying dementia groups from NC’s. The classification performance reached when classifying AD from LBD did not reach the same level. Further research with the aim of developing methods with a higher sensitivity to the different brain changes going on in AD and LBD are needed.