Talk Session 1: Functional Connectivity
10:15am-11:00am
1.1 Recurrent edge-convolutional neural network for resting-state fMRI
Lebo Wang, Kaiming Li, & Xiaoping Hu
Individual variability in static functional connectivity (FC) has been used as a connectomic fingerprinting to differentiate subjects1. However, static FC derives temporal correlation in pairs as spatial pattern, where spatiotemporal features have not been fully explored. Recurrent neural networks (RNNs) have been implemented to model dynamics of fMRI data2. Though better performance was achieved with the fully-connected operation for spatial features, the underlying functional and structural organization of brains was ignored. In this study, we introduced the recurrent edge-convolutional architecture to model the spatial coactivation pattern and dynamics of the fMRI data.
1.2 Connectivity changes after cognitive training in young and aged rats
Luis Colon-Perez, Sean Turner, Katelyn Lubke, Marcelo Febo, & Sara Burke
Changes in large-scale neural connectivity are a hallmark of brain aging that have been linked to cognitive decline. Recent behavioral models for probing the integrity of inter-regional communication along with advances in functional MRI offer a unique opportunity to study functional connectivity in conjunction with quantification of the neurobiological mechanisms that underlie alterations in large-scale network communication. In this study, we determined how cognitive training on an object-place paired association (OPPA) task, which requires interactions between prefrontal, medial temporal and subcortical structures, altered functional connectivity in young (4 mo, n = 5) and aged (24 mo, n = 5) Fischer 344 x Brown Norway F1 hybrid rats. A resting state fMRI dataset was collected in a 11.1 Tesla Bruker system. All rats were scanned for three sessions: before cognitive training, after two weeks of training on the OPPA task in which both young and aged rats were not performing above an 80% criterion (second session), and after four weeks of OPPA training in which the young rats, but not old, were performing at criterion (third session). Time series fMRI signals were extracted from 150 different region of interest (ROI) based on an atlas-guided seed location. The correlation values were organized a graph and thresholded for each subject to create matrices with equal densities (e.g the top 15% correlation values). Young rats did not show differences in global node strengths between the three sessions; however, the aged rats showed an increased in node strength connectivity after the first training session at high node strength values. The areas involved in increasing the node strength values in the aged group were: anterior cingulate, cortical, striatal. Particularly the rich club increased after the first session in the aged group and was maintained during the third session. This suggests an engagement in learning as rats aged in a subnetwork (comprised of anterior cingulate, striatal area, somatosensory, motor, insular, and motor cortex areas) that is not engaged in learning of young rodents.
1.3 Long-range temporal correlations in EEG as a marker of treatment response for infantile spasms
Rachel J. Smith, Hernando C. Ombao, Shaun A. Hussain, Daniel W. Shrey, & Beth A. Lopour
Infantile spasms (IS) is an epileptic encephalopathy that occurs within the first year of life and is characterized by seizures causing abrupt muscle spasms and characteristic electroencephalographic (EEG) patterns. Because IS occurs during critical neurodevelopmental periods, good prognosis depends on prompt successful treatment. To aid in the prediction and assessment of treatment response, we investigated whether a computational EEG measurement of the control of neural synchrony could act as an objective marker of disease burden for IS. We used Detrended Fluctuation Analysis to measure the strength of long-range temporal correlations (LRTCs) in 21 IS patients before treatment and after two weeks of treatment and in 21 control subjects. The strength of LRTCs was significantly lower in pre-treatment subjects than controls, and the two groups could be classified with 92% accuracy using a support vector machine. Successful treatment was associated with a larger increase in LRTCs relative to those in which spasms persisted. This demonstrated that LRTCs changed significantly over the course of two weeks, and the change correlated with treatment response. Because the median response time for treatment is only 2 days, we then established methods to study the continuous timecourse of this change. We developed a moving-block bootstrap (MBB) method for EEG data to calculate time-varying estimates of the strength and confidence intervals of LRTCs. We show that MBB is biased by artifacts in the data but is robust to discontinuities in the time series. We then demonstrate that MBB enables tracking of sleep-wake changes and within-subject comparisons of pre- and post-treatment measurements. These methods will be used to capture the timecourse of changes in temporal structure that correspond to successful treatment outcomes in long-term EEG recorded immediately after treatment initiation. Early changes may predict positive treatment response, informing clinical decisions that will improve patient outcomes.
Talk Session 2: Multivariate Functional MRI in Humans
11:15am-12:15pm
2.1 Decoding item-specific information in visual short-term memory from the hippocampal DG/CA3 subfield using Inverted Encoding Model
Weizhen Xie, Marcus Cappiello, Michael Yassa, Edward Ester, Gopikrishna Deshpande, & Weiwei Zhang
Human memory does not always retain accurate mental representations that precisely correspond to the exceedingly rich contents in natural vision. This functional limit can be attributed to a reduction in precision of internal representations from visual perception to visual short-term memory (VSTM). The mechanism underlying this bottleneck in representational precision remains a topic of controversy. One class of theories attributes VSTM precision to neural noise in sustained neural activities that support VSTM retention. In contrast, another class of theories maintains that VSTM retention and mnemonic precision are supported by dissociable and independent neural mechanisms. For example, the level of neural noise in sustained neural activities for VSTM, which manifests as mnemonic precision of VSTM, may be determined by hippocampal pattern separation, a computational process that orthogonalizes similar memories into non-overlapping representations. To test this hippocampal pattern separation hypothesis, the present study adopted Harrison and Tong’s (2009) orientation VSTM paradigm with high-resolution fMRI. Using the inverted decoding model, we decoded item-specific information from the hippocampal dentate gyrus (DG) and CA3 subfield, a brain region previously implicated in pattern separation, during the delay interval of the VSTM task. In contrast, item-specific information could not be reliably decoded from the hippocampal CA1subfield or the amygdala. A whole-brain searchlight analysis revealed some additional areas in occipital, posterior parietal, and prefrontal cortices that carry item-specific information, replicating some previous findings. Furthermore, Granger causality analyses identified a feedback projection from the hippocampal DG/CA3 to visual cortices during the delay interval, potentially linking hippocampal pattern separation to sensory reactivation of precise representation. Overall, these findings support a novel hippocampal pattern separation hypothesis for mnemonic precision, which is central to the ongoing debate on the nature of the limits in VSTM.
2.2 Examining the role of higher-order visual networks in children’s overgeneralization of conditioned fear
Dana E. Glenn, Megan A. K. Peters, Nathan A. Fox, Daniel S. Pine, & Kalina J. Michalska
Children with Social Reticence (SR), or temperamental shyness, display elevated threat sensitivity and may sometimes inaccurately classify safe stimuli as threatening. This implies important differences in how threat-stimuli are represented in fear circuitry in children with high vs low SR. However, little is known about whether such overgeneralization might stem from a perceptual source. We use Representational Similarity Analysis1 (RSA) in conjunction with functional magnetic resonance imaging (fMRI) to examine the similarity of neural representations of threat- vs safe-stimuli in fear and perceptual neurocircuitry among children classified as high vs low SR. Participants were longitudinally assessed on SR at ages 2, 3, 4, 5, and 7 years. At 11-13 years, 43 children (M=13.38 ±0.62 yrs., 44.2%F) underwent fear conditioning and extinction. One month later, participants returned for fMRI scanning during an extinction recall task in which they viewed “blends” of the CS- and CS+ (Figure 1). We test the hypothesis that high SR youth would show greater overgeneralization to blended cues during extinction recall. We used RSA to compare multi-voxel patterns of activity in the Inferior Temporal (IT) Cortex, a higher-order visual area, in response to CS+, CS- and blends. RSA uses correlations of fMRI voxel activity to quantify the similarity of pairs of neural representations via construction of a Representational Dissimilarity Matrix (Figure 2). We predict that children with low SR will show better differentiation in IT than children with high SR, indicating a perceptual rather than cognitive source for the overgeneralization between CS+ and CS- due to SR.
2.3 Metacognition to accelerate reinforcement learning
Aurelio Cortese, Hakwan Lau, & Mitsuo Kawato
While artificial intelligence algorithms still show severe limitations in real-world scenarios, brains seem to effortlessly solve complex, high dimensional problems from a small sample. One recent model proposed that higher order cognition, e.g. metacognition, may subserve a clear adaptive function. By reducing the dimensionality of the brain representational space(s) and retrieving low-dimensional manifolds, cognition could foster efficient decision-making and accelerate reinforcement learning. Nevertheless, virtually all previous efforts to understand reinforcement learning in the human brain have used tasks with low-dimensional, highly-controlled environments. We developed a novel paradigm whereby participants’ own brain activity measured with fMRI was used in real time to implicitly define hidden task states via a machine learning algorithm (decoder). Due to the large number of relevant voxels, exploring all possible task states would lead to a combinatorial explosion, making sequential searches impossible. This approach allowed us to intrinsically study reinforcement learning in the brain by emulating what the brain faces at all times: how to relate a given action and outcome when degrees-of-freedom are huge. Even in the absence of any physical cue, participants were able to extract low-dimensional manifolds of brain activity to make reward-maximizing choices. We found that metacognition manifested as confidence helped the brain discover such task states through parallel searches. As learning progressed, confidence and reward-prediction errors were increasingly synchronized within prefrontal cortico-basal ganglia loops. The brain can solve complex, high dimensional problems from a small sample without prior information because it can combine metacognition with reinforcement learning mechanisms.
2.4 Sensory preconditioning without conscious awareness: a decoded neurofeedback study
Mouslim Cherkaoui, Cody A. Cushing, Mitsuo Kawato, & Hakwan Lau
It has been shown that when one of two previously-paired neutral items starts to subsequently predict reward, subjects treat the other neutral item as if it is similarly attractive despite the lack of direct evidence linking this item to reward. This phenomenon, known as sensory preconditioning, is often thought to reflect model-based learning, as the formed stimulus association does not reflect reinforced action outcomes. However, there is also evidence suggesting stimulus associations can be formed unconsciously. We tested this hypothesis using machine-learning methods combined with closed-loop ‘online’ fMRI in 5 subjects. Using decoded neurofeedback (DecNef), we formed an unconscious association between a consciously viewed dot motion display and an unconscious image category representation. By analyzing online fMRI data while participants view a dot motion display through a MVPA classifier, feedback was given based on the likelihood that BOLD activation in ventral temporal cortex represented an unrelated target image category (DecNef target), which was never consciously seen or presented. After 3 days of feedback, participants performed a betting task with feedback through which they learn the consciously viewed dot motion display from the feedback sessions results in a significant financial loss. Following this, participants completed another betting task, without feedback, in which they are presented with a critical decision between two previously unseen objects: the DecNef target and a neutral control. All participants chose the neutral control over the DecNef target, indicating successful unconscious preconditioning. These results suggest associations can be formed and conditioned outside of conscious awareness, and that consciousness is not necessary for model-based learning.
Talk Session 3: Structural MRI, Multimodal Imaging, & Other Methodologies
2:15pm-3:00pm
3.1 High temperature superconductor sensors for magnetoencephalography
Shane Cybart
We are developing next-generation Magnetoenceptahalography (MEG) magnetic imaging instrumentation for the human brain. MEG is a non-invasive neural imaging technique that directly measures the magnetic signal due to neuronal activation with high temporal resolution and spatial localization accuracy. MEG has been routinely used in localizing seizure foci in patients with epilepsy and for pre-surgical localization of brain functions. The magnetic field sensors used for commercial MEG systems are based on Superconducting QUantum Interference Devices (SQUID). SQUID-based sensors have the highest combination of sensitivity and bandwidth of any sensor known, but require costly and cumbersome refrigeration for cooling to cryogenic temperatures (4.2 K). The design and materials for these sensors have not changed in nearly three decades and there has been little progress in MEG hardware. The high cost has hindered widespread use of MEG at the clinical level and there are only about 30 of these systems in the United States (two are in the UC system at UCSD and UCSF). We are working on reinventing SQUID MEG, to make it affordable, and easier to use while at the same time improving the sensitivity and resolution. Our approach uses a new type of high temperature superconductor (HTS) SQUID by modifying the ceramic material on the nanoscale using a finely focused beam of helium ions. This results in sensors with very low magnetic flux noise. Furthermore, unlike prior HTS technologies our new technique is relatively simple, inexpensive, high yield, and scalable to wafers for large scale sensor production.
3.2 Deep learning based prognosis for glioblastoma and traumatic brain injury using multi-modal MRI
Bir Bhanu & Padmaja Jonnalagedda
We are working with our collaborators on developing machine learning methods for brain cancer and brain injury using multimodal MRI data DWI-ADC, T2, T1, T1 contrast and others. Glioblastoma multiforme (GBM) is highly common brain tumor in adults, reporting high mortality. Histopathological diagnosis is the primary form of diagnosis. Recent studies have drawn attention to genetic biomarkers in assessing the prognosis and response to treatment. Studies have shown that mutation of IDH1 has a positive correlation with survival probability and MGMT methylation indicates better response to therapy. Research has shown that 19/20 co-gain is a positive prognostic indicator. This research focuses on assessing if imaging features exist that exhibit correlation with concurrent gain of chromosomes 19 and 20 (i.e., 19/20 co-gain). In our preliminary work, the tumor segmentation from MRI images is done using non-linear optimization using particle swarm intelligence technique. The features to be analyzed include tumor location, density, diffusion etc. A random forest classifier is used to classify between mutated and control patients to assess which of the aforementioned modalities (or a combination of them) can best detect 10/20 co-gain and thus provide the prognosis. In addition, imaging biomarkers are being investigated using advanced statistical and machine learning methods. Mild traumatic brain injury (mTBI) is a serious public health problem and major cause of death and disability in the United States, contributing to about 30% of all injury deaths and affecting millions of people. Even a “mild” injury of the brain can result in significant physiological damage and cognitive deficits. TBIs occurs due to events such as sports, military (blast), automobile accidents, assaults, and falls. There are many short and long-term symptoms associated with mTBI including loss of memory, loss of reasoning, neuropsychiatric alterations including decrements in social interactions. However the challenges of current TBI diagnosis is that the symptoms of the patients and signals in the medical imaging are very subtle and they are easily missed by the health care professionals. We attempt to build a framework for using multi-modal and behavioral data to catch subtle indicators leading to prediction or proactive approaches to mTBI. In our preliminary work we have tested the approach using a novel rodent model of repeated mTBI dataset of T2 weighted images (T2WI). The experimental results demonstrate that the fusion of both contextual and visual textural features outperforms other state-of-the-art approaches.
3.3 The role of biological sex on medial temporal lobe structures in healthy and pathological aging: evidence from magnetic resonance imaging (MRI)
Laura M. Ezama-Foronda, Niels Janssen, & Ernesto Pereda De Pablo
A surprising aspect of Alzheimer's Disease (AD) is the higher prevalence of disease in women than in men. Recent research has suggested that this difference does not simply reflect the fact that men die sooner, but there may be genetic differences between the sexes that impact disease probability. If such were the case, biological sex may be a biomarker in the AD. In this study, our interest is focused on several structures that are part of the Medial Temporal Lobe (MTL) related to AD. We will use techniques, such as Magnetic Resonance Imaging (MRI), to analyze how structure and function in MTL areas differs between men and women in various age groups and health status. First, we will examine the structure and function of MTL structures in young healthy men and women. Second, we will examine MTL regions in older healthy men and women. Third, we will study older male and female patients with preclinical symptoms of AD (i.e., “mild cognitive impairment”). The outcome of this project will clarify how biological sex affects MTL structures in normal and pathological aging, and may provide motivation for including sex in AD diagnosis and treatment. Our working hypothesis is that biological sex impacts structure and function in MTL structures during normal and pathological aging. Our basic experimental design will examine the factors Sex, Age and their interactions. The main objective is to clarify the role of biological sex in healthy and pathological aging. Finding positive evidence for differences in hippocampal subfields of brain structure and function between men and women has important consequences for AD disease diagnosis, where biological sex may be included in the diagnosis scheme. In addition, finding that the diseases affects men and women differently may lead to tailored therapies that depend on biological sex.