Borja Ayerdi

Random forest active learning for AAA thrombus segmentation in computed tomography angiography images

Image segmentation of 3D Computed Tomography Angiography (CTA) is affected by a variety of noise conditions that may render ineffective image segmentation procedures that have been developed and validated on a collection of training CTA data when applied on new CTA data. The approach followed in this paper to tackle this problem is to provide an Active Learning based interactive image segmentation system which will allow quick volume segmentation, with minimal intervention of a human operator. Image segmentation is achieved by a Random forest (RF) classifier applied on a set of image features extracted from each voxel and its neighborhood. An initial set of labeled voxels is required to start the process, training an initial RF. The most uncertain unlabeled voxels are shown to the human operator to select some of them for inclusion in the training set, retraining the RF classifier. The approach is applied to the segmentation of the thrombus of Abdominal Aortic Aneurysm (AAA) in CTA data (of patients), showing that the CTA volume can be accurately segmented after few iterations requiring a small labeled data sample.

Hybrid extreme rotation forest

This paper proposes the Hybrid Extreme Rotation Forest (HERF), an innovative ensemble learning algorithm for classification problems, combining classical Decision Trees with the recently proposed Extreme Learning Machines (ELM) training of Neural Networks. In the HERF algorithm, training of each individual classifier involves two steps: first computing a randomized data rotation transformation of the training data, second, training the individual classifier on the rotated data. The testing data is subjected to the same transformation as the training data, which is specific for each classifier in the ensemble. Experimental design in this paper involves (a) the comparison of factorization approaches to compute the randomized rotation matrix: the Principal Component Analysis (PCA) and the Quartimax, (b) assessing the effect of data normalization and bootstrapping training data selection, (c) all variants of single and combined ELM and decision trees, including Regularized ELM. This experimental design effectively includes other state-of-the-art ensemble approaches in the comparison, such as Voting ELM and Random Forest. We report extensive results over a collection of machine learning benchmark databases. Ranking the cross-validation results per experimental dataset and classifier tested concludes that HERF significantly improves over the other state-of-the-art ensemble classifier. Besides, we find some other results such as that the data rotation with Quartimax improves over PCA, and the relative insensitivity of the approach to regularization which may be attributable to the de facto regularization performed by the ensemble approach.

Spatially regularized semisupervised Ensembles of Extreme Learning Machines for hyperspectral image segmentation

This paper explores the performance of Ensembles of Extreme Learning Machine classifiers for hyperspectral image classification and segmentation in a semisupervised and spatially regularized process. The approach assumes that we have available only a small training set of labeled samples, which we enrich with a set of guessed labelings on selected samples from the vast pool of unlabeled image pixels. Selection and label guessing is conditioned to an unsupervised classification of the image pixel spectra, and to the spatial proximity to the labeled samples in the image domain. Unlabeled pixels falling in the spatial neighborhood of a labeled training sample, and belonging to the same unsupervised class, acquire its label. Unsupervised classification can be performed by any clustering technique, in this paper we have resorted to the classical K-means. The classifier built from the enriched training dataset is applied to the entire hyperspectral image. Finally, we perform a spatial regularization of the classification label image, maximizing a rather general prior smoothness criterion, by the selection of the most frequent class in each pixel neighborhood. This paper reports experiments with homogeneous ensembles of ELM, rELM, and OP-ELM classifiers, including a sensitivity analysis over the ensemble size and the number of hidden nodes. Computational experiments on four well known benchmarking hyperspectral images give state-of-the-art results.

Hyperspectral image nonlinear unmixing and reconstruction by ELM regression ensemble

Unmixing is the estimation of hyperspectral image pixels composition, specified as the fractional abundances of the composing materials, achieving image segmentation at sub-pixel resolution. Linear unmixing assumes that pixels are convex combinations of endmember spectra, hence endmember identification is required prior to unmixing processes. In our approach to non-linear unmixing by Extreme Learning Machine (ELM) regression ensembles, we do not need to perform endmember identification, which is implicit in the non-linear transformation. Instead we provide estimates of the fractional abundances of predefined material classes, which have been characterized by pure pixels extracted from the image according to available ground truth. In this paper, we introduce a formal discussion of the convergence properties of ELM regression ensembles that endorses the empirical results. The analysis shows them to converge to the exact regression value when the number of components of the ensemble grows, provided that the output is the average of the individual outputs. Besides, the proposed approach allows for a general validation procedure based on the reconstruction error over the entire hyperspectral image. Reconstruction error can be estimated using the mapping from fractional abundances to reconstructed spectra, also achieved by ELM regression ensembles. Therefore, validation can be carried out independently of training data, which can be used completely for model construction. Experimental results on well known benchmark images show that the approach has big advantage over state-of-the-art unmixing approaches.

Discrimination between Alzheimer’s Disease and Late Onset Bipolar Disorder Using Multivariate Analysis

Background Late onset bipolar disorder (LOBD) is often difficult to distinguish from degenerative dementias, such as Alzheimer disease (AD), due to comorbidities and common cognitive symptoms. Moreover, LOBD prevalence in the elder population is not negligible and it is increasing. Both pathologies share pathophysiological neuroinflammation features. Improvements in differential diagnosis of LOBD and AD will help to select the best personalized treatment. Objective The aim of this study is to assess the relative significance of clinical observations, neuropsychological tests, and specific blood plasma biomarkers (inflammatory and neurotrophic), separately and combined, in the differential diagnosis of LOBD versus AD. It was carried out evaluating the accuracy achieved by classification-based computer-aided diagnosis (CAD) systems based on these variables. Materials A sample of healthy controls (HC) (n = 26), AD patients (n = 37), and LOBD patients (n = 32) was recruited at the Alava University Hospital. Clinical observations, neuropsychological tests, and plasma biomarkers were measured at recruitment time. Methods We applied multivariate machine learning classification methods to discriminate subjects from HC, AD, and LOBD populations in the study. We analyzed, for each classification contrast, feature sets combining clinical observations, neuropsychological measures, and biological markers, including inflammation biomarkers. Furthermore, we analyzed reduced feature sets containing variables with significative differences determined by a Welch’s t-test. Furthermore, a battery of classifier architectures were applied, encompassing linear and non-linear Support Vector Machines (SVM), Random Forests (RF), Classification and regression trees (CART), and their performance was evaluated in a leave-one-out (LOO) cross-validation scheme. Post hoc analysis of Gini index in CART classifiers provided a measure of each variable importance. Results Welch’s t-test found one biomarker (Malondialdehyde) with significative differences (p < 0.001) in LOBD vs. AD contrast. Classification results with the best features are as follows: discrimination of HC vs. AD patients reaches accuracy 97.21% and AUC 98.17%. Discrimination of LOBD vs. AD patients reaches accuracy 90.26% and AUC 89.57%. Discrimination of HC vs LOBD patients achieves accuracy 95.76% and AUC 88.46%. Conclusion It is feasible to build CAD systems for differential diagnosis of LOBD and AD on the basis of a reduced set of clinical variables. Clinical observations provide the greatest discrimination. Neuropsychological tests are improved by the addition of biomarkers, and both contribute significantly to improve the overall predictive performance.

Active Learning with Bootstrapped Dendritic Classifier applied to medical image segmentation

We perform the segmentation of medical images following an Active Learning approach that allows quick interactive segmentation minimizing the requirements for intervention of the human operator. The basic classifier is the Bootstrapped Dendritic Classifier (BDC), which combine the output of an ensemble of weak Dendritic Classifiers by majority voting. Weak Dendritic Classifiers are trained on bootstrapped samples of the train data setting a limit on the number of dendrites. We validate the approach on the segmentation of the thrombus in 3D Computed Tomography Angiography (CTA) data of Abdominal Aortic Aneurysm (AAA) patients simulating the human oracle by the provided ground truth. The generalization results in terms of accuracy and true positive ratio of the classification of the entire volume by the classifier trained on one slice confirm that the approach is worth its consideration for clinical practice.

Hyperspectral Image Analysis by Spectral–Spatial Processing and Anticipative Hybrid Extreme Rotation Forest Classification

Recent classification-oriented proposals to thematic maps building from hyperspectral images have used both semisupervised approaches and spatial information for correction of spectral classification. Semisupervised approaches enrich the training data set adding similar samples to each class, whereas spatial correction is based on the natural assumption of thematic class spatial compactness. In this paper, we propose and validate the following innovations: 1) a new spectral classifier, which is called anticipative hybrid extreme rotation forest (AHERF); 2) a spatial-spectral semisupervised approach; and 3) a final spatial classification correction step. The novel heterogeneous ensemble learning approach AHERF starts with a model selection phase, using a small subsample of the training data, in order to define a ranking-based selection probability distribution of the classifier architectures that will be used in the ensemble, so that the architecture best adapted to the data domain will be used more frequently to train individual classifiers in the ensemble. After this initial phase, AHERF trains a heterogeneous ensemble applying random rotations to bootstrapped samples of the remaining training data, aiming to obtain diversified and data-domain adapted individual classifiers. The natural assumption that spatially close pixels will most likely have highly correlated values is exploited in two phases of the process pipeline. First, semisupervised label assignment is supported by spectral similarity and spatial proximity. Unsupervised spectral similarity is detected by latent class discovery. In this paper, we use a clustering algorithm (i.e., k-means). Second, maximizing class spatial compactness removes classification errors that appear as speckle noise in the classification image. The whole approach aims to use minimal sets of labeled pixels for training, which we call the seed training data set. Testing results are computed over the entire image ground truth. For comparison, we provide results in several steps: 1) of classification by AHERF and competing classifiers built by semisupervised training and 2) after spatial correction. We validate the approach on several conventional benchmarking images, achieving results which are comparable with state-of-the-art approaches.

Risk Factors for Emergency Department Short Time Readmission in Stratified Population

Background. Emergency department (ED) readmissions are considered an indicator of healthcare quality that is particularly relevant in older adults. The primary objective of this study was to identify key factors for predicting patients returning to the ED within 30 days of being discharged. Methods. We analysed patients who attended our ED in June 2014, stratified into four groups based on the Kaiser pyramid. We collected data on more than 100 variables per case including demographic and clinical characteristics and drug treatments. We identified the variables with the highest discriminating power to predict ED readmission and constructed classifiers using machine learning methods to provide predictions. Results. Classifier performance distinguishing between patients who were and were not readmitted (within 30 days), in terms of average accuracy (AC). The variables with the greatest discriminating power were age, comorbidity, reasons for consultation, social factors, and drug treatments. Conclusions. It is possible to predict readmissions in stratified groups with high accuracy and to identify the most important factors influencing the event. Therefore, it will be possible to develop interventions to improve the quality of care provided to ED patients.

Active learning of Hybrid Extreme Rotation Forests for CTA image segmentation

This paper proposes a Hybrid Extreme Rotation Forest (HERF) classifier for segmentation of 3D Computed Tomography Angiography (CTA) following an Active Learning (AL) approach. The HERF is an ensemble of classifiers composed of Extreme Learning Machines (ELM) and Decision Trees. Training of the HERF includes optimal rotation of random partitions of the feature set aimed to increase diversity. AL follows an strategy of optimal sample selection in order to minimize the number of data samples needed to obtain a required accuracy degree. AL is pertinent for interactive learning processes where a human operator is required to select training samples to enhance the classifier in an iterative process, therefore labeling samples for training may be a time consuming and expensive process. CTA image segmentation is one of such processes, due to the variability in CTA images which hinders the generalization of classifiers trained on one dataset to new datasets. Following an AL strategy, the human operator is presented with a visual selection of pixels whose labeling would be most informative for the classifier. After adding those labeled pixels to the training data, the classifier is retrained. This iteration is repeated until image segmentation quality meets the required level. The approach is applied to the segmentation of the thrombus in CTA data of Abdominal Aortic Aneurysm (AAA) patients, showing that the structures of interest in CTA volume can be accurately segmented after a few iterations using a small data sample.

Eigenanatomy on Fractional Anisotropy Imaging Provides White Matter Anatomical Features Discriminating Between Alzheimer’s Disease and Late Onset Bipolar Disorder

BACKGROUND Late Onset Bipolar Disorder (LOBD) is the arousal of Bipolar Disorder (BD) at old age (>60) without any previous history of disorders. LOBD is often difficult to distinguish from degenerative dementias, such as Alzheimer Disease (AD), due to comorbidities and common cognitive symptoms. Moreover, LOBD prevalence is increasing due to population aging. Biomarkers extracted from blood plasma are not discriminant because both pathologies share pathophysiological features related to neuroinflammation, therefore we look for anatomical features highly correlated with blood biomarkers that allow accurate diagnosis prediction. This may shed some light on the basic biological mechanisms leading to one or another disease. Moreover, accurate diagnosis is needed to select the best personalized treatment. OBJECTIVE We look for white matter features which are correlated with blood plasma biomarkers (inflammatory and neurotrophic) discriminating LOBD from AD. MATERIALS A sample of healthy controls (HC) (n=19), AD patients (n=35), and BD patients (n=24) has been recruited at the Alava University Hospital. Plasma biomarkers have been obtained at recruitment time. Diffusion weighted (DWI) magnetic resonance imaging (MRI) are obtained for each subject. METHODS DWI is preprocessed to obtain diffusion tensor imaging (DTI) data, which is reduced to fractional anisotropy (FA) data. In the selection phase, eigenanatomy finds FA eigenvolumes maximally correlated with plasma biomarkers by partial sparse canonical correlation analysis (PSCCAN). In the analysis phase, we take the eigenvolume projection coefficients as the classification features, carrying out cross-validation of support vector machine (SVM) to obtain discrimination power of each biomarker effects. The John Hopkins Universtiy white matter atlas is used to provide anatomical localizations of the detected feature clusters. RESULTS Classification results show that one specific biomarker of oxidative stress (malondialdehyde MDA) gives the best classification performance ( accuracy 85%, F-score 86%, sensitivity, and specificity 87%, ) in the discrimination of AD and LOBD. Discriminating features appear to be localized in the posterior limb of the internal capsule and superior corona radiata. CONCLUSION It is feasible to support contrast diagnosis among LOBD and AD by means of predictive classifiers based on eigenanatomy features computed from FA imaging correlated to plasma biomarkers. In addition, white matter eigenanatomy localizations offer some new avenues to assess the differential pathophysiology of LOBD and AD.