Hierarchical manifold learning for regional image analysis

A Preprocessing Manifold Learning Strategy Based on t-Distributed Stochastic Neighbor Embedding

In machine learning and data analysis, dimensionality reduction and high-dimensional data visualization can be accomplished by manifold learning using a t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm. We significantly improve this manifold learning scheme by introducing a preprocessing strategy for the t-SNE algorithm. In our preprocessing, we exploit Laplacian eigenmaps to reduce the high-dimensional data first, which can aggregate each data cluster and reduce the Kullback-Leibler divergence (KLD) remarkably. Moreover, the k-nearest-neighbor (KNN) algorithm is also involved in our preprocessing to enhance the visualization performance and reduce the computation and space complexity. We compare the performance of our strategy with that of the standard t-SNE on the MNIST dataset. The experiment results show that our strategy exhibits a stronger ability to separate different clusters as well as keep data of the same kind much closer to each other. Moreover, the KLD can be reduced by about 30% at the cost of increasing the complexity in terms of runtime by only 1-2%.

Hierarchical Fully Convolutional Network for Joint Atrophy Localization and Alzheimer's Disease Diagnosis Using Structural MRI

Structural magnetic resonance imaging (sMRI) has been widely used for computer-aided diagnosis of neurodegenerative disorders, e.g., Alzheimer's disease (AD), due to its sensitivity to morphological changes caused by brain atrophy. Recently, a few deep learning methods (e.g., convolutional neural networks, CNNs) have been proposed to learn task-oriented features from sMRI for AD diagnosis, and achieved superior performance than the conventional learning-based methods using hand-crafted features. However, these existing CNN-based methods still require the pre-determination of informative locations in sMRI. That is, the stage of discriminative atrophy localization is isolated to the latter stages of feature extraction and classifier construction. In this paper, we propose a hierarchical fully convolutional network (H-FCN) to automatically identify discriminative local patches and regions in the whole brain sMRI, upon which multi-scale feature representations are then jointly learned and fused to construct hierarchical classification models for AD diagnosis. Our proposed H-FCN method was evaluated on a large cohort of subjects from two independent datasets (i.e., ADNI-1 and ADNI-2), demonstrating good performance on joint discriminative atrophy localization and brain disease diagnosis.

Learning and combining image neighborhoods using random forests for neonatal brain disease classification

It is challenging to characterize and classify normal and abnormal brain development during early childhood. To reduce the complexity of heterogeneous data population, manifold learning techniques are increasingly applied, which find a low-dimensional representation of the data, while preserving all relevant information. The neighborhood definition used for constructing manifold representations of the population is crucial for preserving the similarity structure and it is highly application dependent. The recently proposed neighborhood approximation forests learn a neighborhood structure in a dataset based on a user-defined distance. We propose a framework to learn multiple pairwise distances in a population of brain images and to combine them in an unsupervised manner optimally in a manifold learning step. Unlike other methods that only use a univariate distance measure, our method allows for a natural combination of multiple distances from heterogeneous sources. As a result, it yields a representation of the population that preserves the multiple distances. Furthermore, our method also selects the most predictive features associated with the distances. We evaluate our method in neonatal magnetic resonance images of three groups (term controls, patients affected by intrauterine growth restriction and mild isolated ventriculomegaly). We show that combining multiple distances related to the condition improves the overall characterization and classification of the three clinical groups compared to the use of single distances and classical unsupervised manifold learning.

Unsupervised Learning for Robust Respiratory Signal Estimation From X-Ray Fluoroscopy

Respiratory signals are required for image gating and motion compensation in minimally invasive interventions. In X-ray fluoroscopy, extraction of a respiratory signal can be challenging due to characteristics of interventional imaging, in particular injection of contrast agent and automatic exposure control. We present a novel method for respiratory signal extraction based on dimensionality reduction that can tolerate these events. Images are divided into patches of multiple sizes. Low-dimensional embeddings are generated for each patch using illumination-invariant kernel PCA. Patches with respiratory information are selected automatically by agglomerative clustering. The signals from this respiratory cluster are combined robustly to a single respiratory signal. In the experiments, we evaluate our method on a variety of scenarios. If the diaphragm is visible, we track its superior-inferior motion as ground truth. Our method has a correlation coefficient of more than 91% with the ground truth irrespective of whether or not contrast agent injection or automatic exposure control occur. Additionally, we show that very similar signals are estimated from biplane sequences and from sequences without visible diaphragm. Since all these cases are handled automatically, the method is robust enough to be considered for use in a clinical setting.

Computing group cardinality constraint solutions for logistic regression problems

We derive an algorithm to directly solve logistic regression based on cardinality constraint, group sparsity and use it to classify intra-subject MRI sequences (e.g. cine MRIs) of healthy from diseased subjects. Group cardinality constraint models are often applied to medical images in order to avoid overfitting of the classifier to the training data. Solutions within these models are generally determined by relaxing the cardinality constraint to a weighted feature selection scheme. However, these solutions relate to the original sparse problem only under specific assumptions, which generally do not hold for medical image applications. In addition, inferring clinical meaning from features weighted by a classifier is an ongoing topic of discussion. Avoiding weighing features, we propose to directly solve the group cardinality constraint logistic regression problem by generalizing the Penalty Decomposition method. To do so, we assume that an intra-subject series of images represents repeated samples of the same disease patterns. We model this assumption by combining series of measurements created by a feature across time into a single group. Our algorithm then derives a solution within that model by decoupling the minimization of the logistic regression function from enforcing the group sparsity constraint. The minimum to the smooth and convex logistic regression problem is determined via gradient descent while we derive a closed form solution for finding a sparse approximation of that minimum. We apply our method to cine MRI of 38 healthy controls and 44 adult patients that received reconstructive surgery of Tetralogy of Fallot (TOF) during infancy. Our method correctly identifies regions impacted by TOF and generally obtains statistically significant higher classification accuracy than alternative solutions to this model, i.e., ones relaxing group cardinality constraints.

Echocardiography without electrocardiogram using nonlinear dimensionality reduction methods

PURPOSE

The aim of this study is to evaluate the efficiency of a new automatic image processing technique, based on nonlinear dimensionality reduction (NLDR) to separate a cardiac cycle and also detect end-diastole (ED) (cardiac cycle start) and end-systole (ES) frames on an echocardiography system without using ECG.

METHODS

Isometric feature mapping (Isomap) and locally linear embeddings (LLE) are the most popular NLDR algorithms. First, Isomap algorithm is applied on recorded echocardiography images. By this approach, the nonlinear embedded information in sequential images is represented in a two-dimensional manifold and each image is characterized by a symbol on the constructed manifold. Cyclicity analysis of the resultant manifold, which is derived from the cyclic nature of the heart motion, is used to perform cardiac cycle length estimation. Then, LLE algorithm is applied on extracted left ventricle (LV) echocardiography images of one cardiac cycle. Finally, the relationship between consecutive symbols of the resultant manifold by the LLE algorithm, which is based on LV volume changes, is used to estimate ED (cycle start) and ES frames. The proposed algorithms are quantitatively compared to those obtained by a highly experienced echocardiographer from ECG as a reference in 20 healthy volunteers and 12 subjects with pathology.

RESULTS

Mean difference in cardiac cycle length, ED, and ES frame estimation between our method and ECG detection by the experienced echocardiographer is approximately 7, 17, and 17 ms (0.4, 1, and 1 frame), respectively.

CONCLUSION

The proposed image-based method, based on NLDR, can be used as a useful tool for estimation of cardiac cycle length, ED and ES frames in echocardiography systems, with good agreement to ECG assessment by an experienced echocardiographer in routine clinical evaluation.

Local manifold learning for multiatlas segmentation: application to hippocampal segmentation in healthy population and Alzheimer's disease

AIMS

Automated hippocampal segmentation is an important issue in many neuroscience studies.

METHODS

We presented and evaluated a novel segmentation method that utilized a manifold learning technique under the multiatlas-based segmentation scenario. A manifold representation of local patches for each voxel was achieved by applying an Isomap algorithm, which can then be used to obtain spatially local weights of atlases for label fusion. The obtained atlas weights potentially depended on all pairwise similarities of the population, which is in contrast to most existing label fusion methods that only rely on similarities between the target image and the atlases. The performance of the proposed method was evaluated for hippocampal segmentation and compared with two representative local weighted label fusion methods, that is, local majority voting and local weighted inverse distance voting, on an in-house dataset of 28 healthy adolescents (age range: 10-17 years) and two ADNI datasets of 100 participants (age range: 60-89 years). We also implemented hippocampal volumetric analysis and evaluated segmentation performance using atlases from a different dataset.

RESULTS

The median Dice similarities obtained by our proposed method were approximately 0.90 for healthy subjects and above 0.88 for two mixed diagnostic groups of ADNI subjects.

CONCLUSION

The experimental results demonstrated that the proposed method could obtain consistent and significant improvements over label fusion strategies that are implemented in the original space.

PCA-derived respiratory motion surrogates from X-ray angiograms for percutaneous coronary interventions

Purpose

Intraoperative coronary motion modeling with motion surrogates enables prospective motion prediction in X-ray angiograms (XA) for percutaneous coronary interventions. The motion of coronary arteries is mainly affected by patients breathing and heartbeat. Purpose of our work is therefore to extract coronary motion surrogates that are related to respiratory and cardiac motion. In particular, we focus on respiratory motion surrogates extraction in this paper.

Methods

We propose a fast automatic method for extracting patient-specific respiratory motion surrogate from cardiac XA. The method starts with an image preprocessing step to remove all tubular and curvilinear structures from XA images, such as vessels and guiding catheters, followed by principal component analysis on pixel intensities. The respiratory motion surrogate of an XA image is then obtained by projecting its vessel-removed image onto the first principal component.

Results

This breathing motion surrogate was demonstrated to get high correlation with ground truth diaphragm motion (correlation coefficient over 0.9 on average). In comparison with other related methods, the method we developed did not show significant difference (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p>0.05$$\end{document}p>0.05), but did improve robustness and run faster on monoplane and biplane data in retrospective and prospective scenarios.

Conclusions

we developed and evaluated a method in extraction of respiratory motion surrogate from interventional X-ray images that is easy to implement and runs in real time and thus allows extracting respiratory motion surrogates during interventions.

Regional manifold learning for disease classification

While manifold learning from images itself has become widely used in medical image analysis, the accuracy of existing implementations suffers from viewing each image as a single data point. To address this issue, we parcellate images into regions and then separately learn the manifold for each region. We use the regional manifolds as low-dimensional descriptors of high-dimensional morphological image features, which are then fed into a classifier to identify regions affected by disease. We produce a single ensemble decision for each scan by the weighted combination of these regional classification results. Each weight is determined by the regional accuracy of detecting the disease. When applied to cardiac magnetic resonance imaging of 50 normal controls and 50 patients with reconstructive surgery of Tetralogy of Fallot, our method achieves significantly better classification accuracy than approaches learning a single manifold across the entire image domain.