Diego Castillo-Barnes

Ensemble of random forests One vs. Rest classifiers for MCI and AD prediction using ANOVA cortical and subcortical feature selection and partial least squares

BACKGROUND Alzheimers disease (AD) is the most common cause of dementia in the elderly and affects approximately 30 million individuals worldwide. Mild cognitive impairment (MCI) is very frequently a prodromal phase of AD, and existing studies have suggested that people with MCI tend to progress to AD at a rate of about 10-15% per year. However, the ability of clinicians and machine learning systems to predict AD based on MRI biomarkers at an early stage is still a challenging problem that can have a great impact in improving treatments. METHOD The proposed system, developed by the SiPBA-UGR team for this challenge, is based on feature standardization, ANOVA feature selection, partial least squares feature dimension reduction and an ensemble of One vs. Rest random forest classifiers. With the aim of improving its performance when discriminating healthy controls (HC) from MCI, a second binary classification level was introduced that reconsiders the HC and MCI predictions of the first level. RESULTS The system was trained and evaluated on an ADNI datasets that consist of T1-weighted MRI morphological measurements from HC, stable MCI, converter MCI and AD subjects. The proposed system yields a 56.25% classification score on the test subset which consists of 160 real subjects. COMPARISON WITH EXISTING METHOD(S) The classifier yielded the best performance when compared to: (i) One vs. One (OvO), One vs. Rest (OvR) and error correcting output codes (ECOC) as strategies for reducing the multiclass classification task to multiple binary classification problems, (ii) support vector machines, gradient boosting classifier and random forest as base binary classifiers, and (iii) bagging ensemble learning. CONCLUSIONS A robust method has been proposed for the international challenge on MCI prediction based on MRI data. The system yielded the second best performance during the competition with an accuracy rate of 56.25% when evaluated on the real subjects of the test set.

A semi-supervised learning approach for model selection based on class-hypothesis testing

Abstract This paper deals with the topic of learning from unlabeled or noisy-labeled data in the context of a classification problem. In the classification problem the outcome yields one of a discrete set of values thus, assumptions on them could be established to obtain the most likely prediction model at the training stage . In this paper, a novel case-based model selection method is proposed, which combines hypothesis testing from a discrete set of expected outcomes and feature extraction within a cross-validated classification stage. This wrapper-type procedure acts on fully-observable variables under hypothesis-testing and improves the classification accuracy on the test set, or keeps its performance at least at the level of the statistical classifier. The model selection strategy in the cross validation loop allows building an ensemble classifier that could improve the performance of any expert and intelligence system, particularly on small sample-size datasets. Experiments were carried out on several databases yielding a clear improvement on the baseline, i.e., SPECT dataset A c c = 86.35 ± 1.51 , with S e n = 91.10 ± 2.77 , and S p e = 81.11 ± 1.61 . In addition, the CV error estimate for the classifier under our approach was found to be an almost unbiased estimate (as the baseline approach) of the true error that the classifier would incur on independent data.

A 3D Convolutional Neural Network Approach for the Diagnosis of Parkinson’s Disease

Parkinsonism is the second most common neurodegenerative disease, originated by a dopamine decrease in the striatum. Single Photon Emission Computed Tomography (SPECT) images acquired using the DaTSCAN drug are a widely extended tool in the diagnosis of Parkinson’s Disease (PD), since they can measure the amount of dopamine transporters in the striatum. Many automatic systems have been developed to aid in the diagnosis of PD, using traditional feature extraction methods. In this paper, we propose a novel system based on three-dimensional Convolutional Neural Networks (CNNs), that aims to differenciate between PD-affected patients and unaffected subjects. The proposed system achieves up to a 95.5% accuracy and 96.2% sensitivity in the diagnosis of PD.

Functional Brain Imaging Synthesis Based on Image Decomposition and Kernel Modeling: Application to Neurodegenerative Diseases

The rise of neuroimaging in research and clinical practice, together with the development of new machine learning techniques has strongly encouraged the Computer Aided Diagnosis (CAD) of different diseases and disorders. However, these algorithms are often tested in proprietary datasets to which the access is limited and, therefore, a direct comparison between CAD procedures is not possible. Furthermore, the sample size is often small for developing accurate machine learning methods. Multi-center initiatives are currently a very useful, although limited, tool in the recruitment of large populations and standardization of CAD evaluation. Conversely, we propose a brain image synthesis procedure intended to generate a new image set that share characteristics with an original one. Our system focuses on nuclear imaging modalities such as PET or SPECT brain images. We analyze the dataset by applying PCA to the original dataset, and then model the distribution of samples in the projected eigenbrain space using a Probability Density Function (PDF) estimator. Once the model has been built, we can generate new coordinates on the eigenbrain space belonging to the same class, which can be then projected back to the image space. The system has been evaluated on different functional neuroimaging datasets assessing the: resemblance of the synthetic images with the original ones, the differences between them, their generalization ability and the independence of the synthetic dataset with respect to the original. The synthetic images maintain the differences between groups found at the original dataset, with no significant differences when comparing them to real-world samples. Furthermore, they featured a similar performance and generalization capability to that of the original dataset. These results prove that these images are suitable for standardizing the evaluation of CAD pipelines, and providing data augmentation in machine learning systems -e.g. in deep learning-, or even to train future professionals at medical school.

On a Heavy-Tailed Intensity Normalization of the Parkinson’s Progression Markers Initiative Brain Database

In this work, we normalize the intensity of 40 FP-CIT SPECT images from the Parkinson’s Progression Markers Initiative assuming that the histogram of intensity values follows an \(\alpha \)-stable distribution. Then, we study the normalized images. The interclass separation of the Parkinson’s disease (PD) brain images and the healthy control (HC) are calculated by means of the Mann-Whitney-Wilcoxon U-test. The intensity transformed images present higher inter-class separation according to the estimation of the U-test.

Deep Convolutional Autoencoders vs PCA in a Highly-Unbalanced Parkinson’s Disease Dataset: A DaTSCAN Study

The automated analysis of medical imaging, especially brain imaging, is a challenging high-dimensional task. Computer Aided Diagnosis (CAD) tools often require the images to be spatially normalized and then perform feature extraction to be able to avoid the small sample size problem. However, the spatial normalization often introduces artefacts, especially in functional imaging. Furthermore, variance-based decomposition techniques like PCA, which are extensively used in CAD tools, often perform poorly in highly-unbalanced dataset. To overcome these two problems, we propose a deep Convolutional Autoencoder (CAE) architecture that performs image decomposition -or encoding- in images that were not spatially normalized. A CAD system that used CAE for feature extraction and a Support Vector Machine Classifier (SVC) for classification was tested on a strongly imbalanced (5.69/1) Parkinson’s Disease (PD) neuroimaging dataset from the Parkinson’s Progression Markers Initiative (PPMI), achieving more than 93% accuracy in detecting PD with DaTSCAN imaging, and a area under the ROC curve higher than 0.96. This system paves the way for new deep learning decompositions that bypass the common spatial normalization step and are able to extract relevant information in highly-imbalanced datasets.

Classification Improvement for Parkinson’s Disease Diagnosis Using the Gradient Magnitude in DaTSCAN SPECT Images

In this work, we propose a novel imaging preprocessing step based on the use of the gradient magnitude for medical DaTSCAN SPECT images. As Parkinson’s Disease (PD) is characterized by a marked reduction of intensity at striatum area, measuring intensities in this region is considered as a good marker for this neurological disorder. To extend this idea, we have been studying how quick these values decrease. A simple way to do this was using the gradient of each image. Applying Machine Learning algorithms, we have classified the gradient images and obtained an accuracy improvement of almost 2%. These results prove that the gradient magnitude is even a better marker for PD diagnosis and opens the door to new future investigations about this pathology.

Robust Ensemble Classification Methodology for I123-Ioflupane SPECT Images and Multiple Heterogeneous Biomarkers in the Diagnosis of Parkinson's Disease

In last years, several approaches to develop an effective Computer-Aided-Diagnosis (CAD) system for Parkinsons Disease (PD) have been proposed. Most of these methods have focused almost exclusively on brain images through the use of Machine-Learning algorithms suitable to characterize structural or functional patterns. Those patterns provide enough information about the status and/or the progression at intermediate and advanced stages of Parkinsons Disease. Nevertheless this information could be insufficient at early stages of the pathology. The Parkinsons Progression Markers Initiative (PPMI) database includes neurological images along with multiple biomedical tests. This information opens up the possibility of comparing different biomarker classification results. As data come from heterogeneous sources, it is expected that we could include some of these biomarkers in order to obtain new information about the pathology. Based on that idea, this work presents an Ensemble Classification model with Performance Weighting. This proposal has been tested comparing Healthy Control subjects (HC) vs. patients with PD (considering both PD and SWEDD labeled subjects as the same class). This model combines several Support-Vector-Machine (SVM) with linear kernel classifiers for different biomedical group of tests—including CerebroSpinal Fluid (CSF), RNA, and Serum tests—and pre-processed neuroimages features (Voxels-As-Features and a list of defined Morphological Features) from PPMI database subjects. The proposed methodology makes use of all data sources and selects the most discriminant features (mainly from neuroimages). Using this performance-weighted ensemble classification model, classification results up to 96% were obtained.

Automatic Separation of Parkinsonian Patients and Control Subjects Based on the Striatal Morphology

Parkinsonism is the second more common neurological disease and affects around 1%–2% of people over 65 years, being around 20%–24% of them incorrectly diagnosed. The disorder is associated to a progressive loss of dopaminergic neurons of the striatum. Thus, its diagnosis is usually corroborated by analyzing neuroimaging data of this region. In this work, we propose a novel computer system to automatically distinguish between parkinsonian patients and neurologically healthy subjects using \(^{123}\)I-FP-CIT SPECT data, a neuroimaging modality widely used to assist the diagnosis of Parkinsonism. First, the voxels of the striatum were selected using an intensity threshold. These voxels were then projected over the axial plane, resulting in a two-dimensional image with the striatum shape. Subsequently, the size and shape of the left and right sides of the striatum were characterized by 5 features: area, eccentricity, orientation and length of the major and minor axes. Finally, the extracted features were used along with a Support Vector Machine classifier to separate patients and controls. An accuracy rate of 91.53% (\(p<0.001\)) was estimated using a k-fold cross-validation scheme and a database with 189 \(^{123}\)I-FP-CIT SPECT neuroimages. This rate outperformed the ones achieved by previous approaches when using the same data.

A Heavy Tailed Expectation Maximization Hidden Markov Random Field Model with Applications to Segmentation of MRI

A wide range of segmentation approaches assumes that intensity histograms extracted from magnetic resonance images (MRI) have a distribution for each brain tissue that can be modeled by a Gaussian distribution or a mixture of them. Nevertheless, intensity histograms of White Matter and Gray Matter are not symmetric and they exhibit heavy tails. In this work, we present a hidden Markov random field model with expectation maximization (EM-HMRF) modeling the components using the α-stable distribution. The proposed model is a generalization of the widely used EM-HMRF algorithm with Gaussian distributions. We test the α-stable EM-HMRF model in synthetic data and brain MRI data. The proposed methodology presents two main advantages: Firstly, it is more robust to outliers. Secondly, we obtain similar results than using Gaussian when the Gaussian assumption holds. This approach is able to model the spatial dependence between neighboring voxels in tomographic brain MRI.