Machine learning approaches have had some success in predicting conversion to Alzheimer’s Disease (AD) in subjects with mild cognitive impairment (MCI), a less serious condition that nonetheless is a risk factor for AD. Predicting conversion is clinically important as because novel drugs currently being developed require administration early in the disease process to be effective. Traditionally training data are labelled with discrete disease states; which may explain the limited accuracies obtained as labels are noisy due to the difficulty in providing a definitive diagnosis of Alzheimer’s without post-mortem confirmation, and ignore the existence of a continuous spectrum of disease severity. Here, we dispense with discrete training labels and instead predict the loss of brain volume over one year, a quantity that can be repeatably and objectively measured with the boundary shift integral and is strongly correlated with conversion. The method combines MRI and PET image data and cerebrospinal fluid biomarker levels in an Bayesian multi-kernel learning framework. The resulting predicted atrophy separates converting and non-converting MCI subjects with 74.6 % accuracy, which compares well to state of the art methods despite a small training set size.

Classification of brain images is frequently done using kernel based methods, such as the support vector machine. These lend themselves to improvement via multiple kernel learning, where a number of different kernels are linearly combined to integrate different sources of information and increase accuracy. Previous applications made use of a small number of kernels representing different image modalities or kernel functions. Here, the kernels instead represent 83 anatomically meaningful brain regions. To find the optimal combination of kernels and perform classification, we use a Gaussian Process framework to infer the maximum likelihood weights. The resulting formulation successfully combines voxel level features with prior anatomical knowledge. This gives an improvement in classification accuracy of MRI images of Alzheimer's disease patients and healthy controls from the ADNI database to almost 88 %, compared to less than 86% using a single kernel representing the whole brain. Moreover, interpretability of the classifier is also improved, as the optimal kernel weights are sparse and give an indication of the importance of each brain region in separating the two groups.

Posterior cortical atrophy (PCA) is a neurodegenerative syndrome characterized by predominant visual deficits and parieto-occipital atrophy, and is typically associated with Alzheimer's disease (AD) pathology. In AD, assessment of hippocampal atrophy is widely used in diagnosis, research, and clinical trials; its utility in PCA remains unclear. Given the posterior emphasis of PCA, we hypothesized that hippocampal shape measures may give additional group differentiation information compared with whole-hippocampal volume assessments. We investigated hippocampal volume and shape in subjects with PCA (n = 47), typical AD (n = 29), and controls (n = 48). Hippocampi were outlined on MRI scans and their 3D meshes were generated. We compared hippocampal volume and shape between disease groups. Mean adjusted hippocampal volumes were ∼ 8% smaller in PCA subjects (P < 0.001) and ∼ 22% smaller in tAD subject (P < 0.001) compared with controls. Significant inward deformations in the superior hippocampal tail were observed in PCA compared with controls even after adjustment for hippocampal volume. Inward deformations in large areas of the hippocampus were seen in tAD subjects compared with controls and PCA subjects, but only localized shape differences remained after adjusting for hippocampal volume. The shape differences observed, even allowing for volume differences, suggest that PCA and tAD are each associated with different patterns of hippocampal tissue loss that may contribute to the differential range and extent of episodic memory dysfunction in the two groups.

Machine learning has obvious applications to the diagnosis of disease, and for many neurological conditions features extracted from brain images allow classifiers based on neuroimaging biomarkers to provide a useful complement to more traditional diagnostic methods based on symptoms and psychological testing. However the labels used in the training of such systems frequently depend on standard clinical diagnostic methods, meaning they are not completely reliable in many cases. This uncertainty makes the problems this causes hard to study, as it is difficult to measure both the extent of mislabelling and its effect on results. To avoid this problem, we perform classification of gender based on imaging, as this is definitely known for each subject. We then deliberately make known proportions of the training labels incorrect. This allows us to assess the effect of the level of label noise on classification accuracy, and evaluate methods that allow for the mislabelled data. The methods are wrappers using existing well known classifier algorithms. The results indicate that the methods can be significantly effective at realistic levels of noise in the training labels, but care must be taken in choosing which method to apply depending on the level of label noise.

Accurately identifying the patients that have mild cognitive impairment (MCI) who will go on to develop Alzheimer's disease (AD) will become essential as new treatments will require identification of AD patients at earlier stages in the disease process. Most previous work in this area has centred around the same automated techniques used to diagnose AD patients from healthy controls, by coupling high dimensional brain image data or other relevant biomarker data to modern machine learning techniques. Such studies can now distinguish between AD patients and controls as accurately as an experienced clinician. Models trained on patients with AD and control subjects can also distinguish between MCI patients that will convert to AD within a given timeframe (MCI-c) and those that remain stable (MCI-s), although differences between these groups are smaller and thus, the corresponding accuracy is lower. The most common type of classifier used in these studies is the support vector machine, which gives categorical class decisions. In this paper, we introduce Gaussian process (GP) classification to the problem. This fully Bayesian method produces naturally probabilistic predictions, which we show correlate well with the actual chances of converting to AD within 3 years in a population of 96 MCI-s and 47 MCI-c subjects. Furthermore, we show that GPs can integrate multimodal data (in this study volumetric MRI, FDG-PET, cerebrospinal fluid, and APOE genotype with the classification process through the use of a mixed kernel). The GP approach aids combination of different data sources by learning parameters automatically from training data via type-II maximum likelihood, which we compare to a more conventional method based on cross validation and an SVM classifier. When the resulting probabilities from the GP are dichotomised to produce a binary classification, the results for predicting MCI conversion based on the combination of all three types of data show a balanced accuracy of 74%. This is a substantially higher accuracy than could be obtained using any individual modality or using a multikernel SVM, and is competitive with the highest accuracy yet achieved for predicting conversion within three years on the widely used ADNI dataset.

There has been a great deal of recent work making use of multivariate classification techniques such as support vector machines to classify brain images obtained from MRI or PET as healthy or suffering from neurodegenerative disease. In the case of Alzheimer's disease, the results are as accurate as standard clinical tests and could potentially be used in a diagnostic setting. However these techniques give categorical class decisions. Here we show for the first time that Gaussian processes can be applied to structural neuroimaging data to perform classification of Alzheimer's disease subjects in a fully Bayesian framework. This offers advantages such as automatic setting of parameters via type II maximum likelihood and probabilistic predictions that may be useful in a clinical context, while maintaining the same accuracy as a state-of-the-art discriminative classifier applied to the same data.

It is well known that hippocampal atrophy is a marker of the onset of Alzheimer's disease (AD) and as a result hippocampal volumetry has been used in a number of studies to provide early diagnosis of AD and predict conversion of mild cognitive impairment patients to AD. However, rates of atrophy are not uniform across the hippocampus making shape analysis a potentially more accurate biomarker. This study studies the hippocampi from 226 healthy controls, 148 AD patients and 330 MCI patients obtained from T1 weighted structural MRI images from the ADNI database. The hippocampi are anatomically segmented using the MAPS multi-atlas segmentation method, and the resulting binary images are then processed with SPHARM software to decompose their shapes as a weighted sum of spherical harmonic basis functions. The resulting parameterizations are then used as feature vectors in Support Vector Machine (SVM) classification. A wrapper based feature selection method was used as this considers the utility of features in discriminating classes in combination, fully exploiting the multivariate nature of the data and optimizing the selected set of features for the type of classifier that is used. The leave-one-out cross validated accuracy obtained on training data is 88.6% for classifying AD vs controls and 74% for classifying MCI-converters vs MCI-stable with very compact feature sets, showing that this is a highly promising method. There is currently a considerable fall in accuracy on unseen data indicating that the feature selection is sensitive to the data used, however feature ensemble methods may overcome this.

Hippocampal atrophy is a marker of the onset of Alzheimer's disease (AD) and hippocampal volumetry has been used in a number of studies to provide early diagnosis of AD and predict conversion of Mild Cognitive Impairment (MCI) to AD. However rates of atrophy are not uniform across the hippocampus, making shape a potentially more sensitive biomarker than volume. This study investigated the utility of SPHARM parameterizations of hippocampal shape coupled with a support vector machine (SVM) classifier to predict conversion from MCI to AD. Left and right hippocampi from a total of 330 images of MCI patients from the ADNI database were automatically segmented. The resulting volumes were then processed with to produce a description of each shape as a weighted sum of spherical harmonic basis functions. The weights of each shape were taken as a feature vector for SVM classification. Separate classifiers were built for right and left hippocampal shape features and for a concatenation of the two. For each experiment, two thirds of the subjects were used to build a classifier. The shape features were ranked according to their ability to separate MCI converter and MCI stable subjects, and then a ten fold cross validation was performed for each classifier to find the number of hippocampal features and SVM parameters yielding the best classification results. The resulting classifiers were then applied to an unseen test set comprising the remaining third of the subjects. Performance with features from the right hippocampus only was superior to left only or both together. As AUC is considered a better measure of classification efficacy than accuracy, the parameters yielding the best AUC were used when classifying unseen data. In this case classification accuracy was 68% with a binomial (exact) 95% confidence interval of 58% to 77% and an AUC of 0.74. Details are given in the table. This method shows promise in discriminating between MCI stable and MCI converter subjects. The best classification results are obtained with a somewhat biased classifier, meaning that specificity was substantially better than sensitivity, however the AUC shows that this classifier can potentially be both accurate and balanced.