Alzheimer’s Disease Classification with Multi-Modal Image Data
 
Authors:
Vikas Singh, PhD, University of Wisconsin-Madison; Chris Hinrichs, MS; Sterling C. Johnson, PhD; Guofan Xu, MD, PhD
 
Background:
Alzheimer’s disease (AD) is an irreversible neurodegenerative disease that affects over five million people in the United States, with almost 400,000 new cases identified each year. Currently, there is no cure for the AD, though several promising drugs are under investigation. As better treatments become available, it will be important to identify candidate patients for treatment as early as possible, in order to prevent the neuro-degenerative change that is characteristic of AD. Measures derived from magnetic resonance (MR) structural imaging (e.g., gray matter density), have been proposed as a surrogate marker for the early diagnosis of AD. Similarly, studies using fluorodeoxyglucose-positron emission tomography (FDG PET) images have shown that impaired metabolism is also a marker for AD. Further, performance on cognitive tests, family history, and other clinical biomarker data show strong correlations with incidence and onset of AD. A strong emphasis in neuroscience research is to leverage all this information for early identification, so that interventions can be planned and treatments (as they become available) can be adopted. Therefore, the development of machine learning tools to enable a multi-modal image classification pipeline for Alzheimer’s disease is urgently needed – for diagnosis at the level of individual subjects as well as to better understand the progression of the disease.
 
Evaluation:
We make use of a method which is an extension of well-known max-margin kernel learning, such as Support Vector Machines (SVMs). SVMs seek a linear classifier (i.e., a separating hyperplane) to separate the two classes of (positively and negatively) labeled examples, while also finding the widest possible gap, or margin, between them. A more recent development has been the extension of the SVM framework to utilize multiple kernels, where each represents a different view of the same example (e.g., each imaging modality represents a different view of a subject). This framework is known as Multi-Kernel Learning (MKL). In our experiments we observed that a subset of our data contain outliers (i.e., subjects who are abnormal with respect to the clinical group to which they have been designated). This can be due to a variety of factors, including: co-morbidity with other diseases; and early signs of atrophy, which predate cognitive decline or even misdiagnosis of the type of dementia due to the unavailability of gold standard post-mortem diagnoses. Motivated by this, and in order to obtain highly accurate classifiers, we adapted the MKL model to include automated outlier detection, which not only identifies outlier subjects during the training phase, but also reduces their influence on the learned classifier. In summary, the framework allows mitigating the influence of outlier examples on the classifier, while offering the capability to work with multiple imaging modalities simultaneously.
 
Discussion:
We evaluated our multi-modal learning framework on image scans from the ADNI dataset. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) is a landmark research study sponsored by the National Institutes of Health, to determine whether brain imaging can help predict onset and monitor progression of Alzheimer’s disease. The study is ongoing and will cover a total of 800 patients (200 healthy controls, 400 MCI individuals, and 200 mild AD patients). For our experiments, we used MR and PET scans of 159 patients (77 AD and 82 controls) from this dataset. The data also provides a diagnosis for each subject based on clinical evaluations. This was used for training the classifier and for calculating the accuracy of the system. To evaluate our algorithm, we adopted a two-fold approach. First, we evaluated the efficacy of the multi-kernel framework as a classification system, with respect to its accuracy using ROC curves. Second, we evaluated the clinical interpretation of the regions selected as being discriminative by our model.

Figure 1

Figure 1

We first tested the accuracy of our method by comparing the classifier’s output on unseen test examples with their clinical group assignments. By sorting the examples by the classifier’s outputs, and using each example’s output as a separate threshold, we can plot an ROC curve, which compares the rate of false positives with true positives (sensitivity) for each threshold. These curves are shown in Fig.1. The corresponding areas under this curve (AUC) is a representative measure of the accuracy of the method. For these experiments, we found that the AUC for the proposed method was 0.885, when using both MR and FDG-PET modalities. In addition, we also performed a clinical interpretation of the regions corresponding to the learned classifier. These regions are shown in Figs. 2 and 3, corresponding to MR and FDG-PET, respectively. The classifier selected the hippocampus and parahippocampal gyri, as well as other middle-temporal regions in MR, while in FDG-PET the posterior cingulate cortex and parietal lobules bilaterally appear prominently. These regions are known to be strongly affected by AD pathology, and are an encouraging form of validation for our method.

 

Figure 2

Figure 2

 

Figure 3

Figure 3
 
Conclusion:
This study describes a method for Alzheimer’s disease (AD) classification using multi-modal image data. Our experiments have concentrated primarily on AD, but we believe that the basic machinery is general enough to be used without difficulty in classification (or diagnosis) applications where one is faced with a problem of how to combine distinct types of information (i.e., demography data, laboratory test scores, and images). We observed that this is typically accomplished by simply concatenating all available data about a patient into a long feature vector, which is then fed into a standard machine learning framework. This may be sufficient for some applications, but the accuracy may suffer because such an approach assigns equal weight to every modality. The proposed scheme allows easy and seamless combinations of different types of images, and yields good accuracy for AD diagnosis. We believe these ideas will find applications in other image-based computer assisted diagnosis problems.
 
References:
Hinrichs C, Singh V, Mukherjee L, Xu G, Chung MK, Johnson SC. “Spatially Augmented LP Boosting for AD classification.” To appear in Neuroimage.

Sonnenburg S, Ratsch G, Schafer C, Scholkopf B. “Large Scale Multi-kernel Learning.” Journal of Machine Learning Research. 2006;Vol.