Grigori Yourganov

Machine Learning in Medical Imaging

This article will discuss very different ways of using machine learning that may be less familiar, and we will demonstrate through examples the role of these concepts in medical imaging. Although the term machine learning is relatively recent, the ideas of machine learning have been applied to medical imaging for decades, perhaps most notably in the areas of computer-aided diagnosis (CAD) and functional brain mapping. We will not attempt in this brief article to survey the rich literature of this field. Instead our goals will be 1) to acquaint the reader with some modern techniques that are now staples of the machine-learning field and 2) to illustrate how these techniques can be employed in various ways in medical imaging.

Dimensionality Estimation for Optimal Detection of Functional Networks in BOLD fMRI data

Estimation of the intrinsic dimensionality of fMRI data is an important part of data analysis that helps to separate the signal of interest from noise. We have studied multiple methods of dimensionality estimation proposed in the literature and used these estimates to select a subset of principal components that was subsequently processed by linear discriminant analysis (LDA). Using simulated multivariate Gaussian data, we show that the dimensionality that optimizes signal detection (in terms of the receiver operating characteristic (ROC) metric) goes through a transition from many dimensions to a single dimension as a function of the signal-to-noise ratio. This transition happens when the loci of activation are organized into a spatial network and the variance of the networked, task-related signals is high enough for the signal to be easily detected in the data. We show that reproducibility of activation maps is a metric that captures this switch in intrinsic dimensionality. Except for reproducibility, all of the methods of dimensionality estimation we considered failed to capture this transition: optimization of Bayesian evidence, minimum description length, supervised and unsupervised LDA prediction, and Steins unbiased risk estimator. This failure results in sub-optimal ROC performance of LDA in the presence of a spatially distributed network, and may have caused LDA to underperform in many of the reported comparisons in the literature. Using real fMRI data sets, including multi-subject group and within-subject longitudinal analysis we demonstrate the existence of these dimensionality transitions in real data.

Optimizing Preprocessing and Analysis Pipelines for Single-Subject fMRI: 2. Interactions with ICA, PCA, Task Contrast and Inter-Subject Heterogeneity

A variety of preprocessing techniques are available to correct subject-dependant artifacts in fMRI, caused by head motion and physiological noise. Although it has been established that the chosen preprocessing steps (or “pipeline”) may significantly affect fMRI results, it is not well understood how preprocessing choices interact with other parts of the fMRI experimental design. In this study, we examine how two experimental factors interact with preprocessing: between-subject heterogeneity, and strength of task contrast. Two levels of cognitive contrast were examined in an fMRI adaptation of the Trail-Making Test, with data from young, healthy adults. The importance of standard preprocessing with motion correction, physiological noise correction, motion parameter regression and temporal detrending were examined for the two task contrasts. We also tested subspace estimation using Principal Component Analysis (PCA), and Independent Component Analysis (ICA). Results were obtained for Penalized Discriminant Analysis, and model performance quantified with reproducibility (R) and prediction metrics (P). Simulation methods were also used to test for potential biases from individual-subject optimization. Our results demonstrate that (1) individual pipeline optimization is not significantly more biased than fixed preprocessing. In addition, (2) when applying a fixed pipeline across all subjects, the task contrast significantly affects pipeline performance; in particular, the effects of PCA and ICA models vary with contrast, and are not by themselves optimal preprocessing steps. Also, (3) selecting the optimal pipeline for each subject improves within-subject (P,R) and between-subject overlap, with the weaker cognitive contrast being more sensitive to pipeline optimization. These results demonstrate that sensitivity of fMRI results is influenced not only by preprocessing choices, but also by interactions with other experimental design factors. This paper outlines a quantitative procedure to denoise data that would otherwise be discarded due to artifact; this is particularly relevant for weak signal contrasts in single-subject, small-sample and clinical datasets.

Multivariate Connectome-Based Symptom Mapping in Post-Stroke Patients: Networks Supporting Language and Speech.

Language processing relies on a widespread network of brain regions. Univariate post-stroke lesion-behavior mapping is a particularly potent method to study brain–language relationships. However, it is a concern that this method may overlook structural disconnections to seemingly spared regions and may fail to adjudicate between regions that subserve different processes but share the same vascular perfusion bed. For these reasons, more refined structural brain mapping techniques may improve the accuracy of detecting brain networks supporting language. In this study, we applied a predictive multivariate framework to investigate the relationship between language deficits in human participants with chronic aphasia and the topological distribution of structural brain damage, defined as post-stroke necrosis or cortical disconnection. We analyzed lesion maps as well as structural connectome measures of whole-brain neural network integrity to predict clinically applicable language scores from the Western Aphasia Battery (WAB). Out-of-sample prediction accuracy was comparable for both types of analyses, which revealed spatially distinct, albeit overlapping, networks of cortical regions implicated in specific aspects of speech functioning. Importantly, all WAB scores could be predicted at better-than-chance level from the connections between gray-matter regions spared by the lesion. Connectome-based analysis highlighted the role of connectivity of the temporoparietal junction as a multimodal area crucial for language tasks. Our results support that connectome-based approaches are an important complement to necrotic lesion-based approaches and should be used in combination with lesion mapping to fully elucidate whether structurally damaged or structurally disconnected regions relate to aphasic impairment and its recovery. SIGNIFICANCE STATEMENT We present a novel multivariate approach of predicting post-stroke impairment of speech and language from the integrity of the connectome. We compare it with multivariate prediction of speech and language scores from lesion maps, using cross-validation framework and a large (n = 90) database of behavioral and neuroimaging data from individuals with post-stroke aphasia. Connectome-based analysis was similar to lesion-based analysis in terms of predictive accuracy and provided additional details about the importance of specific connections (in particular, between parietal and posterior temporal areas) for preserving speech functions. Our results suggest that multivariate predictive analysis of the connectome is a useful complement to multivariate lesion analysis, being less dependent on the spatial constraints imposed by underlying vasculature.

Comparing classification methods for longitudinal fmri studies

We compare 10 methods of classifying fMRI volumes by applying them to data from a longitudinal study of stroke recovery: adaptive Fishers linear and quadratic discriminant; gaussian naive Bayes; support vector machines with linear, quadratic, and radial basis function (RBF) kernels; logistic regression; two novel methods based on pairs of restricted Boltzmann machines (RBM); and K-nearest neighbors. All methods were tested on three binary classification tasks, and their out-of-sample classification accuracies are compared. The relative performance of the methods varies considerably across subjects and classification tasks. The best overall performers were adaptive quadratic discriminant, support vector machines with RBF kernels, and generatively trained pairs of RBMs.

PHYCAA: data-driven measurement and removal of physiological noise in BOLD fMRI.

The effects of physiological noise may significantly limit the reproducibility and accuracy of BOLD fMRI. However, physiological noise evidences a complex, undersampled temporal structure and is often non-orthogonal relative to the neuronally-linked BOLD response, which presents a significant challenge for identifying and removing such artifact. This paper presents a multivariate, data-driven method for the characterization and removal of physiological noise in fMRI data, termed PHYCAA (PHYsiological correction using Canonical Autocorrelation Analysis). The method identifies high frequency, autocorrelated physiological noise sources with reproducible spatial structure, using an adaptation of Canonical Correlation Analysis performed in a split-half resampling framework. The technique is able to identify physiological effects with vascular-linked spatial structure, and an intrinsic dimensionality that is task- and subject-dependent. We also demonstrate that increasing dimensionality of such physiological noise is correlated with increasing variability in externally-measured respiratory and cardiac processes. Using PHYCAA as a denoising technique significantly improves simulated signal detection with physiological noise, and real data-driven model prediction and reproducibility, for both block and event-related task designs. This is demonstrated compared to no physiological noise correction, and to the widely used RETROICOR (Glover et al., 2000) physiological denoising algorithm, which uses externally measured cardiac and respiration signals.

Dimensionality of brain networks linked to life-long individual differences in self-control

The ability to delay gratification in childhood has been linked to positive outcomes in adolescence and adulthood. Here we examine a subsample of participants from a seminal longitudinal study of self-control throughout a subject’s lifespan. Self control, first studied in children at age 4, is now reexamined 40 years later, on a task that required control over the contents of working memory. We examine whether patterns of brain activation on this task can reliably distinguish participants with consistently low and high self-control abilities (low vs. high delayers). We find that low delayers recruit significantly higher-dimensional neural networks when performing the task compared to high delayers. High delayers are also more homogeneous as a group in their neural patterns compared to low delayers. From these brain patterns we can predict with 71% accuracy, whether a participant is a high or low delayer. The present results suggest that dimensionality of neural networks is a biological predictor of self-control abilities.

Pattern classification of fMRI data: Applications for analysis of spatially distributed cortical networks

The field of fMRI data analysis is rapidly growing in sophistication, particularly in the domain of multivariate pattern classification. However, the interaction between the properties of the analytical model and the parameters of the BOLD signal (e.g. signal magnitude, temporal variance and functional connectivity) is still an open problem. We addressed this problem by evaluating a set of pattern classification algorithms on simulated and experimental block-design fMRI data. The set of classifiers consisted of linear and quadratic discriminants, linear support vector machine, and linear and nonlinear Gaussian naive Bayes classifiers. For linear discriminant, we used two methods of regularization: principal component analysis, and ridge regularization. The classifiers were used (1) to classify the volumes according to the behavioral task that was performed by the subject, and (2) to construct spatial maps that indicated the relative contribution of each voxel to classification. Our evaluation metrics were: (1) accuracy of out-of-sample classification and (2) reproducibility of spatial maps. In simulated data sets, we performed an additional evaluation of spatial maps with ROC analysis. We varied the magnitude, temporal variance and connectivity of simulated fMRI signal and identified the optimal classifier for each simulated environment. Overall, the best performers were linear and quadratic discriminants (operating on principal components of the data matrix) and, in some rare situations, a nonlinear Gaussian naïve Bayes classifier. The results from the simulated data were supported by within-subject analysis of experimental fMRI data, collected in a study of aging. This is the first study that systematically characterizes interactions between analysis model and signal parameters (such as magnitude, variance and correlation) on the performance of pattern classifiers for fMRI.

Is the preference of natural versus man-made scenes driven by bottom-up processing of the visual features of nature?

Previous research has shown that viewing images of nature scenes can have a beneficial effect on memory, attention, and mood. In this study, we aimed to determine whether the preference of natural versus man-made scenes is driven by bottom–up processing of the low-level visual features of nature. We used participants’ ratings of perceived naturalness as well as esthetic preference for 307 images with varied natural and urban content. We then quantified 10 low-level image features for each image (a combination of spatial and color properties). These features were used to predict esthetic preference in the images, as well as to decompose perceived naturalness to its predictable (modeled by the low-level visual features) and non-modeled aspects. Interactions of these separate aspects of naturalness with the time it took to make a preference judgment showed that naturalness based on low-level features related more to preference when the judgment was faster (bottom–up). On the other hand, perceived naturalness that was not modeled by low-level features was related more to preference when the judgment was slower. A quadratic discriminant classification analysis showed how relevant each aspect of naturalness (modeled and non-modeled) was to predicting preference ratings, as well as the image features on their own. Finally, we compared the effect of color-related and structure-related modeled naturalness, and the remaining unmodeled naturalness in predicting esthetic preference. In summary, bottom–up (color and spatial) properties of natural images captured by our features and the non-modeled naturalness are important to esthetic judgments of natural and man-made scenes, with each predicting unique variance.

Revealing the dual streams of speech processing

Significance Relatively recently, the concept of dual route neural architecture, where dorsal and ventral brain regions process information synergistically, has been applied to study of speech processing. Although a large body of work has investigated these streams in relation to human speech processing, there is little consensus regarding specific cortical regions implicated. Relying on extensive behavioral and neuroimaging data from a large sample of stroke survivors, we used a data-driven approach to localize regions crucial for motor–phonological and lexical–semantic aspects of speech processing. Results revealed distinct anatomical boundaries between a dorsal frontoparietal stream supporting a form-to-articulation pathway and a ventral temporal–frontal stream supporting a form-to-meaning pathway. This study shows clear division between two processing routes underlying human speech. Several dual route models of human speech processing have been proposed suggesting a large-scale anatomical division between cortical regions that support motor–phonological aspects vs. lexical–semantic aspects of speech processing. However, to date, there is no complete agreement on what areas subserve each route or the nature of interactions across these routes that enables human speech processing. Relying on an extensive behavioral and neuroimaging assessment of a large sample of stroke survivors, we used a data-driven approach using principal components analysis of lesion-symptom mapping to identify brain regions crucial for performance on clusters of behavioral tasks without a priori separation into task types. Distinct anatomical boundaries were revealed between a dorsal frontoparietal stream and a ventral temporal–frontal stream associated with separate components. Collapsing over the tasks primarily supported by these streams, we characterize the dorsal stream as a form-to-articulation pathway and the ventral stream as a form-to-meaning pathway. This characterization of the division in the data reflects both the overlap between tasks supported by the two streams as well as the observation that there is a bias for phonological production tasks supported by the dorsal stream and lexical–semantic comprehension tasks supported by the ventral stream. As such, our findings show a division between two processing routes that underlie human speech processing and provide an empirical foundation for studying potential computational differences that distinguish between the two routes.