Moving from Detection to Pre-detection of Alzheimer’s Disease from MRI Data


Alzheimer’s Disease (AD) is the most common form of dementia, affecting approximately 10% of individuals under 65 years of age, with the prevalence doubling every 5 years up to age 80, above which the prevalence exceeds 40%. Currently diagnosis of AD is largely based on the examination of clinical history and tests such as MMSE (Mini–mental state examination) and PAL (Paired Associates Learning). However many present studies have highlighted the inaccuracies and limitations of such tests. Thus medical officers are now moving to the more accurate neuroimaging data (Magnetic Resonance Imaging- MRI) based diagnosis for these types of diseases where brain atrophy transpires. However it is a considerable challenge to analyse large numbers of images manually to get the most accurate diagnosis at present.

During the recent years, there have been many studies on automatic diagnosis of AD using different methods. The focus of most of these studies has been on detection of AD from neuroimaging data. However, recognizing symptoms early (Pre-detection) is crucial as disease modifying drugs will be most effective if administered early in the course of disease before irreversible brain damage occurs. Therefore we believe that, it is of high importance to utilize automated techniques for the purpose for pre-detection of AD symptoms from such data. The design of this study is twofold.

  1. Firstly we attempt to verify whether the methods used for detection of AD in previous studies can be used successfully for pre-detection, where the automated identification will be more challenging due to the weaker presence of the symptoms.
  2. Secondly, if those tests don’t yield successful results, we plan to propose a deep learning based technique for this purpose.

For our initial experiments, we used MRI data made available as part of the AD Neuroimaging Initiative.  This data set consist of 755 patients in each one of the three classes (AD, Mild Cognitive Impairment – MCI, Healthy Aged Subjects – HC), for a total of 2,428 scans. A sample group of 69 were chosen, from this data set, for the preliminary study. Samples contained a series of longitudinal volumetric T1 MRI scans of 46 mild-moderate Alzheimer’s subjects and 23 controls. Recent studies have shown that ventricular measures may be an early and potentially pre-symptomatic disease biomarker, hence, in our initial tests we aimed to detect the cerebral ventricle volume from MRI scans for atrophy. Initially Pre-processing, Region of Interest (ROI) selection and edge detection algorithms were applied using MATLAB to ease the detection of atrophy. (Fig 1).

q

w

e-copy

Fig 1: Pre-processing, edge detection and ROI selection using MATLAB

As Support Vector Machines (SVM) have been used extensively for the de

tection studies, this technique was used for our initial tests, where we attempted to train and test a SVM with linear kernels which was implemented using MATLAB’s svmtrain and svmclassify functions. As the results of this test, we observed that while 13.13% were misclassified for AD cases, 1.46% were wrongly classified for healthy controls (Fig 2- Highlighted areas). Therefore it is apparent that the previously utilized SVM method is not ideal to detect symptoms of mild to moderate AD cases (Pre-detection stage). Thus we plan to use, the Convolutional Neural Networks (CNN) for brain image classification with multiple output nodes for the next stage of this study.

capture

Fig 2: SVM Classification of AD and Healthy Controls

The reasons for the selection of CNNs are as follows. Since we have to classify the data sets into more than 2 classes, it is not ideal to use SVM, as it is not a good option for multiple class classifications. Further, as we have to deal with a huge number of images in these sorts of studies for both testing and training, SVM is not suited for this task. Studies have shown that CNN is more suited for relatively large data sets and multiple class classifications as compared with the performance of SVM for similar tasks.

If the proposed pre-detection mechanism for AD is successful and fully implemented, it will significantly improve the accuracy of early detection of AD and be substantially beneficial for the treatment.