Classifying Alzheimer's disease, Lewy body dementia, and normal controls using 3D texture analysis in magnetic resonance images

Dementia is an evolving challenge in society and early intervention is important. The ability to distinguish between different dementia and non-dementia early in course may be essential for successful patient care. Magnetic resonance (MR) imaging may aid as a noninvasive method to increase prediction accuracy. In this work we explored the use of two different 3D local binary pattern (LBP) texture features extracted from T1 MR images of the brain combined with a random forest classifier in an attempt to discern patients with Alzheimer's disease (AD), Lewy body dementia (LBD), and normal controls (NC). Analysis were conducted in areas with white matter lesions (WML) and normal appearing white matter (NAWM). We also calculated correlations between texture features and cognition measured by mini mental state examination (MMSE) controlling for age. Additionally, two different methods for handling the imbalanced data problem were tested, namely cost-sensitive classification and resampling of the data using the synthetic minority oversampling technique (SMOTE). Four different classification tasks were extensively tested, a three-class problem: AD vs. LBD vs. NC, a two-class problem: NC vs. AD, a two-class problem NC vs. LBD, and a two-class problem: AD vs. LBD. Results from 10 folds nested cross validation are reported as mean accuracy, precision, and recall with standard deviation in brackets. The two-class problems NC vs. AD and NC vs. LBD, show encouraging results with total accuracy of 0.97 (0.07) and 0.97 (0.06) respectively. The three-class problem and the two-class problem AD vs. LBD are not equally encouraging but shows higher accuracy than clinical diagnosis with a total accuracy of 0.79 (0.07) and 0.79 (0.15) respectively. Possible explanations may be that the AD- and LBD group are too similar concerning LBP texture analysis and that the LBD group is too small. Most of the texture features calculated for the AD subjects in the NAWM region were significantly correlated with cognition. Together with the positive classification results from the NAWM region this may suggest that the NAWM region is an important area for studying AD. Both cost-sensitive classification and resampling using SMOTE proved useful and improved the results considerably in many of the tests.