Prediction of Alzheimer disease using improved MMSE ensemble regressor based on magnetic resonance images
Cognitive scores are the most common measures in diagnosing Alzheimer’s disease which are measured clinically. These scores are mostly useful in late and severe stages of disease which symptoms of the disease are appeared. Nowadays, it is obvious that onset of the disease can be even decades before...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/57591/1/FK%202015%2077RR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Cognitive scores are the most common measures in diagnosing Alzheimer’s disease which are measured clinically. These scores are mostly useful in late and severe stages of disease which symptoms of the disease are appeared. Nowadays, it is obvious that onset of the disease can be even decades before manifestation of the symptoms and it can be revealed by investigating the brain structures. Early prognosing of Alzheimer’s
disease by analyzing brain MR images and inspecting effect of it on brain structures is a hard task. Moreover, predicting severity of disease based on the cognitive scores is a more challenging process especially for future prediction by using the anatomical parameters in the past. One of major problems is high dimensionality of anatomical feature space which must be reduced to a small feature set of discriminative ones and another issue is to relate them to the cognitive scores in the future. This thesis addresses these problems and investigates in the relationships between AD progression and brain degenerations. Brain MR Images of Alzheimer’s Disease NeuroImaging (ADNI) dataset are used in the thesis. A total of 108 subjects who pass the imaging process at four successive time scans of screening, 12th month, 24th month and 36th month are selected. 30 subjects from Normal Controls (NC), 30 from Alzheimer’s disease (AD) holders, 30 subject with Mild Cognitive Impairments (MCI) and 18 converters, all convert at 36th month, from MCI to AD are included in the dataset. Brain MR Images are analysed by the established Freesurfer algorithms to extract the volumetric and thickness features of brain structures in all four time scans. These features are used as raw data in the rest of the thesis. The thesis has four major objectives. First,
discriminative features which vary significantly during the disease monitoring period are identified according to the cognitive scores. Next, regarding to the ordered nature of cognitive scores it aims to find those features that impose smaller error to the cognitive scores as predicted output values. These two objectives are going to solve high dimensionality issue. To tackle on relationship between cognitive scores in future and
anatomical features, third objective is proposed to find a relationship between the selected anatomical features throughout the monitoring period and MMSE scores at the
end of period or 36th month. It is obvious that using anatomical feature values at 36th month to predict the MMSE scores at the same time is clinically unworthy. Fourth objective is to overcome this shortcoming and relate the anatomical feature values at the 36th month to those of the screening, 12th month and 24th month. To achieve the first objective, an evolutionary hypothesis test is proposed to reduce the feature size and chose those ones that their variation during the 36 months of screening is significant and was not stable in the duration. Additionally, they must differ significantly according to the cognitive scores. A minimal set of feature who passed the above criteria and can differentiate all of cognitive score pairs is selected by using a genetic search algorithm. Chernoff bound as upper bound of Bayes error for class separability is computed for evaluating the feature selection method. A reduction from 69.1 to 50.2 is achieved for the proposed evolutionary hypothesis test. In the proposed feature selection algorithm, the ordered nature of cognitive scores or ranks and the amount of error value that can be imposed by any feature over any rank are never considered. So, a rank based feature selection algorithm is proposed to address these issues. It assigns three measures to any pair of feature and rank. These three measures are sorted in each rank and truncated based on a threshold of their derivatives. Those features that are kept in all three truncated feature sets are chosen as final selected features which are 10 features. Chernoff bound decreases again from 50.2 to 46.3 by using the rank based feature selection algorithm. As noted in the third objective, these selected features are used to predict the MMSE scores at the 36th month of screening. Four various core regressors are used including multilayer perceptron regressor,general regression neural network, support vector regressor and relevant vector
regressor. Each of the core regressors participate in a boosting algorithm and then, a bulk of 40 regressors participate in designing final ensemble regressor. To this end, the feature space must be clustered into some small perfect hyperspaces. Each hyperspace is assumed as perfect hyperspace if at least three of the regressors can predict perfectly all data pattern in it. Averaging method is adopted for predicting MMSE scores in any
cluster. Mean square error value of 0.0112 and correlation coefficient of 0.9556 reveal competence of the proposed method. Predicting MMSE scores of 36th month by using the anatomical features of the same time is not clinically beneficial. To address it and accomplish the fourth objective, some ensemble regressors are proposed to predict anatomical features of 36th month or long term features by using their short term counterparts from start of screening up to the 24th month. The same proposed ensemble regression method is used in designing these regressors. Mean square errors range between 0.0064 and 0.0111 and correlation coefficients range between 0.8393 and 0.09355 indicate suitability of proposed algorithm even in predicting other type of features. The really measured long term features in the designed ensemble regressor are replaced by the predicted counterparts to achieve a feasible MMSE ensemble regressor. A mean square error of 0.0213 and correlation coefficient of 0.9350 indicates that the feasible ensemble regressor is a good representative of the one which is designed by the real long term features. |
|---|