Biology subjects

Liew, S.-L.

Publications and source records attributed to Liew, S.-L..

13 recordsLinked to original sources

bioRxiv2026

CALM-VLM: CALIBRATION AND SELECTIVE PREDICTION IN VISION-LANGUAGE MODELS FOR RELIABLE BRAIN MRI CLASSIFICATION

AO_SCPLOWBSTRACTC_SCPLOWRecent advances in vision-language models (VLMs) have demonstrated strong multimodal capabilities for medical image analysis. However, their confidence in diagnostic predictions is often unclear, limiting adoption in clinical settings. We introduce CALM-VLM (CALibration Mechanism for Vision-Language Models), which integrates confidence calibration and selective prediction into a generative 3D VLM. To create CALM-VLM, we fine-tuned the Med3DVLM architecture for Alzheimers disease (AD) and stroke classification as initial test cases. To improve reliability, we incorporated temperature scaling on the VLMs generative outputs. The calibrated model then selectively abstained from predictions when uncertain; this also improved its diagnostic accuracy. Experiments across multi-site MRI datasets, from 10 countries worldwide, show that CALM-VLM improved confidence relative to uncalibrated VLMs. Coverage-adjusted test receiver-operator characteristic curve-area under the curve (ROC-AUC) increased by 5% to 13% for both diagnostic tasks across independent test sets. Our calibrated VLM achieved a test ROC-AUC of 0.951 for AD classification and 0.905 for stroke classification. These findings highlight the importance of calibrated, uncertainty-aware VLMs for trustworthy neuroimaging AI.

neuroscience↗

bioRxiv2023

Data-driven biomarkers outperform theory-based biomarkers in predicting stroke motor outcomes

Chronic motor impairments are a leading cause of disability after stroke. Previous studies have predicted motor outcomes based on the degree of damage to predefined structures in the motor system, such as the corticospinal tract. However, such theory-based approaches may not take full advantage of the information contained in clinical imaging data. The present study uses data-driven approaches to predict chronic motor outcomes after stroke and compares the accuracy of these predictions to previously-identified theory-based biomarkers. Using a cross-validation framework, regression models were trained using lesion masks and motor outcomes data from 789 stroke patients (293 female/496 male) from the ENIGMA Stroke Recovery Working Group (age 64.9{+/-}18.0 years; time since stroke 12.2{+/-}0.2 months; normalised motor score 0.7{+/-}0.5 (range [0,1]). The out-of-sample prediction accuracy of two theory-based biomarkers was assessed: lesion load of the corticospinal tract, and lesion load of multiple descending motor tracts. These theory-based prediction accuracies were compared to the prediction accuracy from three data-driven biomarkers: lesion load of lesion-behaviour maps, lesion load of structural networks associated with lesion-behaviour maps, and measures of regional structural disconnection. In general, data-driven biomarkers had better prediction accuracy - as measured by higher explained variance in chronic motor outcomes - than theory-based biomarkers. Data-driven models of regional structural disconnection performed the best of all models tested (R2 = 0.210, p < 0.001), performing significantly better than predictions using the theory-based biomarkers of lesion load of the corticospinal tract (R2 = 0.132, p< 0.001) and of multiple descending motor tracts (R2 = 0.180, p < 0.001). They also performed slightly, but significantly, better than other data-driven biomarkers including lesion load of lesion-behaviour maps (R2 =0.200, p < 0.001) and lesion load of structural networks associated with lesion-behaviour maps (R2 =0.167, p < 0.001). Ensemble models - combining basic demographic variables like age, sex, and time since stroke - improved prediction accuracy for theory-based and data-driven biomarkers. Finally, combining both theory-based and data-driven biomarkers with demographic variables improved predictions, and the best ensemble model achieved R2 = 0.241, p < 0.001. Overall, these results demonstrate that models that predict chronic motor outcomes using data-driven features, particularly when lesion data is represented in terms of structural disconnection, perform better than models that predict chronic motor outcomes using theory-based features from the motor system. However, combining both theory-based and data-driven models provides the best predictions.

neuroscience↗

bioRxiv2022

Global brain health modulates the impact of lesion damage on post-stroke sensorimotor outcomes

Sensorimotor performance after stroke is strongly related to focal injury measures such as corticospinal tract lesion load. However, the role of global brain health is less clear. Here, we examined the impact of brain age, a measure of neurobiological aging derived from whole brain structural neuroimaging, on sensorimotor outcomes. We hypothesized that stroke lesion damage would result in older brain age, which would in turn be associated with poorer sensorimotor outcomes. We also expected that brain age would mediate the impact of lesion damage on sensorimotor outcomes and that these relationships would be driven by post-stroke secondary atrophy (e.g., strongest in the ipsilesional hemisphere in chronic stroke). We further hypothesized that structural brain resilience, which we define in the context of stroke as the brains ability to maintain its global integrity despite focal lesion damage, would differentiate people with better versus worse outcomes. We analyzed cross-sectional high-resolution brain MRI and outcomes data from 963 people with stroke from 38 cohorts worldwide using robust linear mixed-effects regressions to examine the relationship between sensorimotor behavior, lesion damage, and brain age. We used a mediation analysis to examine whether brain age mediates the impact of lesion damage on stroke outcomes and if associations are driven by ipsilesional measures in chronic ([≥]180 days) stroke. We assessed the impact of brain resilience on sensorimotor outcome using logistic regression with propensity score matching on lesion damage. Stroke lesion damage was associated with older brain age, which in turn was associated with poorer sensorimotor outcomes. Brain age mediated the impact of corticospinal tract lesion load on sensorimotor outcomes most strongly in the ipsilesional hemisphere in chronic stroke. Greater brain resilience, as indexed by younger brain age, explained why people have better versus worse sensorimotor outcomes when lesion damage was fixed. We present novel evidence that global brain health is associated with superior post-stroke sensorimotor outcomes and modifies the impact of focal damage. This relationship appears to be due to post-stroke secondary degeneration. Brain resilience provides insight into why some people have better outcomes after stroke, despite similar amounts of focal injury. Inclusion of imaging-based assessments of global brain health may improve prediction of post-stroke sensorimotor outcomes compared to focal injury measures alone. This investigation is important because it introduces the potential to apply novel therapeutic interventions to prevent or slow brain aging from other fields (e.g., Alzheimers disease) to stroke.

neuroscience↗

bioRxiv2021

Corpus callosal microstructure predicts bimanual motor performance in chronic stroke survivors

AO_SCPLOWBSTRACTC_SCPLOWMicrostructural changes in the corpus callosum are associated with more severe motor impairment in the paretic hand, poor recovery, and general disability. Considering its role in bimanual coordination, we suspected that these microstructural changes across the callosum may also be reflected in the performance of ecologically valid routine bimanual tasks. Thus, the purpose of this study was to determine if callosal microstructure predicts bimanual motor performance in chronic stroke survivors by examining the regions of the corpus callosum connecting both the sensorimotor and non-sensorimotor cortices. We examined the relationship between the fractional anisotropy across the CC and movement times for two self-initiated and self-paced bimanual tasks in 41 chronic stroke survivors. Using publicly available control datasets (n = 52), matched closely for acquisition parameters, we also explored the effect of stroke and age on callosal microstructure. There were two main findings: First, callosal microstructure was significantly associated with bimanual performance in chronic stroke survivors. Notably, a significant relationship was observed not only with the primary sensorimotor regions, but also regions of the premotor/supplementary motor and prefrontal regions. Second, chronic stroke survivors presented with significantly lower mean FA, compared to neurologically intact adults. We conclude that in mild-to-moderate chronic stroke survivors with relatively localized lesions to the motor areas, callosal microstructure can be expected to change in not only the primary sensorimotor region, but also more anteriorly in the secondary motor regions and the genu and is associated with performance on cooperative bimanual tasks. SignificanceA goal of rehabilitation after stroke is to promote the return to pre-stroke levels of upper limb function and use, predominantly characterized by coordinated bimanual activities. In this study, we find that in the chronic phase of stroke, microstructural disorganization within the corpus callosum predicts motor performance on real-world bimanual tasks and lends important insight into the indirect, remote effects of stroke. HO_SCPLOWIGHLIGHTSC_SCPLOWO_LIWe provide initial evidence that corpus callosal microstructure predicts performance on two self-initiated and self-paced bimanual tasks. C_LIO_LIAssociations were strongest for fibers connecting the primary sensorimotor cortices followed by the pre- and supplementary motor, and prefrontal cortices. C_LIO_LISignificant reductions in fractional anisotropy were observed in stroke survivors for all regions of the corpus callosum. C_LI

neuroscience↗

bioRxiv2020

Non-lesioned subcortical brain volumes are associated with post-stroke sensorimotor behavior across 28 cohorts worldwide: An ENIGMA Stroke Recovery study

Background and PurposeUp to two-thirds of stroke survivors experience persistent sensorimotor impairments. Recovery relies on the integrity of spared brain areas to compensate for damaged tissue. Subcortical regions play critical roles in the control and regulation of sensorimotor circuits. The goal of this work is to identify associations between volumes of spared subcortical nuclei and sensorimotor behavior at different timepoints after stroke. MethodsWe pooled high-resolution T1-weighted MRI brain scans and behavioral data in 828 individuals with unilateral stroke from 28 cohorts worldwide. Cross-sectional analyses using linear mixed-effects models related post-stroke sensorimotor behavior to non-lesioned subcortical volumes (Bonferroni-corrected, p<0.004). We tested subacute ([≤]90 days) and chronic ([≥]180 days) stroke subgroups separately, with exploratory analyses in early stroke ([≤]21 days) and across all time. Sub-analyses in chronic stroke were also performed based on class of sensorimotor deficits (impairment, activity limitations) and side of lesioned hemisphere. ResultsWorse sensorimotor behavior was associated with a smaller ipsilesional thalamic volume in both early (n=179; d=0.68) and subacute (n=274, d=0.46) stroke. In chronic stroke (n=404), worse sensorimotor behavior was associated with smaller ipsilesional putamen (d=0.52) and nucleus accumbens (d=0.39) volumes, and a larger ipsilesional lateral ventricle (d=-0.42). Worse chronic sensorimotor impairment specifically (measured by the Fugl-Meyer Assessment; n=256) was associated with smaller ipsilesional putamen (d=0.72) and larger lateral ventricle (d=-0.41) volumes, while several measures of activity limitations (n=116) showed no significant relationships. In the full cohort across all time (n=828), sensorimotor behavior was associated with the volumes of the ipsilesional nucleus accumbens (d=0.23), putamen (d=0.33), thalamus (d=0.33), and lateral ventricle (d=-0.23). ConclusionsWe demonstrate significant relationships between post-stroke sensorimotor behavior and reduced volumes of subcortical gray matter structures that were spared by stroke, which differ by time and class of sensorimotor measure. These findings may provide additional targets for improving post-stroke sensorimotor outcomes.

neuroscience↗

bioRxiv2020

Determinants of Real-Time fMRI Neurofeedback Performance and Improvement: a Machine Learning Mega-Analysis

Real-time fMRI neurofeedback is an increasingly popular neuroimaging technique that allows an individual to gain control over his/her own brain signals, which can lead to improvements in behavior in healthy participants as well as to improvements of clinical symptoms in patient populations. However, a considerably large ratio of participants undergoing neurofeedback training do not learn to control their own brain signals and, consequently, do not benefit from neurofeedback interventions, which limits clinical efficacy of neurofeedback interventions. As neurofeedback success varies between studies and participants, it is important to identify factors that might influence neurofeedback success. Here, for the first time, we employed a big data machine learning approach to investigate the influence of 20 different design-specific (e.g. activity vs. connectivity feedback), region of interest-specific (e.g. cortical vs. subcortical) and subject-specific factors (e.g. age) on neurofeedback performance and improvement in 608 participants from 28 independent experiments. With a classification accuracy of 60% (considerably different from chance level), we identified two factors that significantly influenced neurofeedback performance: Both the inclusion of a pre-training no-feedback run before neurofeedback training and neurofeedback training of patients as compared to healthy participants were associated with better neurofeedback performance. The positive effect of pre-training no-feedback runs on neurofeedback performance might be due to the familiarization of participants with the neurofeedback setup and the mental imagery task before neurofeedback training runs. Better performance of patients as compared to healthy participants might be driven by higher motivation of patients, higher ranges for the regulation of dysfunctional brain signals, or a more extensive piloting of clinical experimental paradigms. Due to the large heterogeneity of our dataset, these findings likely generalize across neurofeedback studies, thus providing guidance for designing more efficient neurofeedback studies specifically for improving clinical neurofeedback-based interventions. To facilitate the development of data-driven recommendations for specific design details and subpopulations the field would benefit from stronger engagement in Open Science and data sharing.

neuroscience↗

bioRxiv2020

Testing a convolutional neural network-based hippocampal segmentation method in a stroke population

As stroke mortality rates decrease, there has been a surge of effort to study post-stroke dementia (PSD) to improve long-term quality of life for stroke survivors. Hippocampal volume may be an important neuroimaging biomarker in post-stroke dementia, as it has been associated with many other forms of dementia. However, studying hippocampal volume using MRI requires hippocampal segmentation. Advances in automated segmentation methods have allowed for studying the hippocampus on a large scale, which is important for robust results in the heterogeneous stroke population. However, most of these automated methods use a single atlas-based approach and may fail in the presence of severe structural abnormalities common in stroke. Hippodeep, a new convolutional neural network-based hippocampal segmentation method, does not rely solely on a single atlas-based approach and thus may be better suited for stroke populations. Here, we compared quality control and the accuracy of segmentations generated by Hippodeep and two well-accepted hippocampal segmentation methods on stroke MRIs (FreeSurfer 6.0 whole hippocampus and FreeSurfer 6.0 sum of hippocampal subfields). Quality control was performed using a stringent protocol for visual inspection of the segmentations, and accuracy was measured as volumetric and spatial comparisons to the manual segmentations. Hippodeep performed significantly better than both FreeSurfer methods in terms of quality control and spatial accuracy. Overall, this study suggests that both Hippodeep and FreeSurfer may be useful for hippocampal segmentation in stroke rehabilitation research, but Hippodeep may be more robust to stroke lesion anatomy.

neuroscience↗

bioRxiv2020

Can we predict real-time fMRI neurofeedback learning success from pre-training brain activity?

Neurofeedback training has been shown to influence behavior in healthy participants as well as to alleviate clinical symptoms in neurological, psychosomatic, and psychiatric patient populations. However, many real-time fMRI neurofeedback studies report large interindividual differences in learning success. The factors that cause this vast variability between participants remain unknown and their identification could enhance treatment success. Thus, here we employed a meta-analytic approach including data from 24 different neurofeedback studies with a total of 401 participants, including 140 patients, to determine whether levels of activity in target brain regions during pre-training functional localizer or no-feedback runs (i.e., self-regulation in the absence of neurofeedback) could predict neurofeedback learning success. We observed a slightly positive correlation between pre-training activity levels during a functional localizer run and neurofeedback learning success, but we were not able to identify common brain-based success predictors across our diverse cohort of studies. Therefore, advances need to be made in finding robust models and measures of general neurofeedback learning, and in increasing the current study database to allow for investigating further factors that might influence neurofeedback learning.

neuroscience↗

bioRxiv2019

Embodiment improves performance on an immersive brain computer interface in head-mounted virtual reality

Brain computer interfaces (BCI) can be used to provide individuals with neurofeedback of their own brain activity and train them to learn how to control their brain activity. Neurofeedback-based BCIs used for motor rehabilitation aim to close the loop between attempted motor commands and sensory feedback by providing supplemental sensory information when individuals successfully establish specific brain patterns. Existing neurofeedback-based BCIs have used a variety of displays to provide feedback, ranging from devices that provide a more immersive and compelling experience (e.g., head-mounted virtual reality (HMD-VR) or CAVE systems) to devices that are considered less immersive (e.g., computer screens). However, it is not clear whether more immersive systems (i.e., HMD-VR) improve neurofeedback performance compared to computer screens, and whether there are individual performance differences in HMD-VR versus screen-based neurofeedback. In this pilot experiment, we compared neurofeedback performance in HMD-VR versus on a computer screen in twelve healthy individuals. We also examined whether individual differences in presence or embodiment correlated with neurofeedback performance in either environment. Participants were asked to control a virtual right arm by imagining right hand movements. Real-time brain activity indicating motor imagery, which was measured via electroencephalography (EEG) as desynchronized sensorimotor rhythms (SMR; 8-24 Hz) in the left motor cortex, drove the movement of the virtual arm towards (increased SMR desynchronization) or away from (decreased SMR desynchronization) targets. Participants performed two blocks of 30 trials, one for each condition (Screen, HMD-VR), with the order of conditions counterbalanced across participants. After completing each block, participants were asked questions relating to their sense of presence and embodiment in each environment. We found that, while participants performance on the neurofeedback-based BCI task was similar between conditions, the participants reported levels of embodiment was significantly different between conditions. Specifically, participants experienced higher levels of embodiment in HMD-VR compared to the computer screen. We further found that reported levels of embodiment positively correlated with neurofeedback performance only in the HMD-VR condition. Overall, these preliminary results suggest that embodiment may improve performance on a neurofeedback-based BCI and that HMD-VR may increase embodiment during a neurofeedback-based BCI task compared to a standard computer screen.

neuroscience↗

bioRxiv2018

High-definition transcranial direct current stimulation over the dorsal premotor cortex reliably modulates motor network excitability

The effectiveness of transcranial direct current stimulation (tDCS) placed over the motor hotspot (thought to represent the primary motor cortex (M1)) to modulate motor network excitability is highly variable. The premotor cortex--particularly the dorsal premotor cortex (PMd)--may be a promising alternative target to more effectively modulate motor excitability, as it influences motor control across multiple pathways, one independent of M1 and one with direct, modulating connections to M1. This double-blind, placebo-controlled study aimed to differentially excite motor and premotor regions using high-definition tDCS (HD-tDCS) with concurrent functional magnetic resonance imaging (fMRI). HD-tDCS applied over either the motor hotspot or the premotor cortex demonstrated high inter-individual variability in changes on cortical motor excitability. However, HD-tDCS over the premotor cortex led to a higher number of responders and greater changes in local fMRI-based complexity than HD-tDCS over the motor hotspot. Furthermore, an analysis of individual motor hotspot anatomical locations revealed that, in more than half of the participants, the motor hotspot is not located over anatomical M1 boundaries, despite using a canonical definition of the motor hotspot. This heterogeneity in stimulation site may contribute to the variability of tDCS results. Altogether, these findings provide new considerations to enhance tDCS reliability.

neuroscience↗

bioRxiv2018

A comparison of automated lesion segmentation approaches for chronic stroke T1-weighted MRI data

Accurate stroke lesion segmentation is a critical step in the neuroimaging processing pipeline to assess the relationship between post-stroke brain structure, function, and behavior. While many multimodal segmentation algorithms have been developed for acute stroke neuroimaging, few are effective with only a single T1-weighted (T1w) anatomical MRI. This is a critical gap because most stroke rehabilitation research relies on a single T1w MRI for defining the lesion. Although several attempts to automate the segmentation of chronic lesions on single-channel T1w MRI have been made, these approaches have not been systematically evaluated on a large dataset. Here, we performed an exhaustive review of the literature and identified one semi- and three fully automated approaches for segmentation of chronic stroke lesions using T1w MRI within the last ten years: Clusterize, Automated Lesion Identification, Gaussian naive Bayes lesion detection, and LINDA. We evaluated each method on a large T1w stroke dataset (N=181) using visual and quantitative methods. LINDA was the most computationally expensive approach, but performed best across the three main evaluation metrics (median values: Dice Coefficient=0.50, Hausdorffs Distance=36.34 mm, and Average Symmetric Surface Distance = 4.97 mm), whereas the Gaussian Bayes method had the highest recall/least false negatives (median=0.80). Segmentation accuracy in all automated methods were influenced by size (small: worst) and stroke territory (brainstem, cerebellum: worst) of the lesion. To facilitate reproducible science, we have made our analysis files publicly available online at https://github.com/npnl/elsa. We hope these findings are informative to future development of T1w lesion segmentation algorithms.

neuroscience↗

bioRxiv2017

The Anatomical Tracings of Lesions After Stroke (ATLAS) Dataset - Release 1.1

Stroke is the leading cause of adult disability worldwide, with up to two-thirds of individuals experiencing long-term disabilities. Large-scale neuroimaging studies have shown promise in identifying robust biomarkers (e.g., measures of brain structure) of long-term stroke recovery following rehabilitation. However, analyzing large rehabilitation-related datasets is problematic due to barriers in accurate stroke lesion segmentation. Manually-traced lesions are currently the gold standard for lesion segmentation on T1-weighted MRIs, but are labor intensive and require anatomical expertise. While algorithms have been developed to automate this process, the results often lack accuracy. Newer algorithms that employ machine-learning techniques are promising, yet these require large training datasets to optimize performance. Here we present ATLAS (Anatomical Tracings of Lesions After Stroke), an open-source dataset of 304 T1-weighted MRIs with manually segmented lesions and metadata. This large, diverse dataset can be used to train and test lesion segmentation algorithms and provides a standardized dataset for comparing the performance of different segmentation methods. We hope ATLAS release 1.1 will be a useful resource to assess and improve the accuracy of current lesion segmentation methods.

neuroscience↗

bioRxiv2017

Distributed cortical structural properties contribute to motor cortical excitability and inhibition

The link between the local structure of the primary motor cortex and motor function has been well documented. However, motor function relies on a network of interconnected brain regions and the link between the structural properties characterizing these distributed brain networks and motor function remains poorly understood. Here, we examined whether distributed patterns of brain structure, extending beyond the primary motor cortex can help classify two forms of motor function: corticospinal excitability and intracortical inhibition. To this effect, we recorded high-resolution structural magnetic resonance imaging scans in 25 healthy volunteers. To measure corticospinal excitability and inhibition in the same volunteers we recorded motor evoked potentials (MEPs) elicited by single-pulse transcranial magnetic stimulation (TMS) and short-interval intracortical inhibition (SICI) in a separate session. Support vector machine (SVM) pattern classification was used to identify distributed multivoxel gray matter areas, which distinguished subjects who had lower and higher MEPs and SICIs. We found that MEP and SICI classification could be predicted based on a widely distributed, largely non-overlapping pattern of voxels in the frontal, parietal, temporal, occipital and cerebellar regions. Thus, structural properties distributed over the brain beyond the primary motor cortex relate to motor function.

neuroscience↗