Learning through language requires the brain to rapidly transform words into meaningful concepts, but how this process unfolds over time remains unclear. Using a combination of functional MRI and electroencephalography (EEG), we tracked brain activity in healthy adults as they processed language. We identified a sequence of brain networks that became active within about 700 milliseconds, progressing from visual word recognition to retrieving meaning, interpreting context, integrating sentence structure and meaning, and ultimately forming a coherent concept. Transitions between these stages involved shared brain regions, particularly the left temporoparietal junction. People with stronger language comprehension showed greater involvement of later processes related to interpreting and integrating information. These findings reveal how different brain networks work together over fractions of a second to transform language into conceptual knowledge.

Figure 1

Spatial and temporal perspectives on language processing. Current neuroimaging approaches examine language processing through separate methodological silos, limiting our understanding of how distributed brain networks coordinate dynamically in real time. a) Spatial perspective: brain systems supporting language comprehension, after��Fedorenko et��al. (2024), showing the distributed architecture of perception of linguistic input (green), core language processing (red), motor planning (navy), and higher-order cognition (purple). Regions are rendered on an inflated left-hemisphere cortical surface from probabilistic functional atlases for the language, multiple demand (light purple), and default mode (dark purple) networks (Lipkin et��al. 2022;��Yeo et��al. 2011), and Harvard-Oxford for the perception and motor areas. While fMRI reveals the spatial organization of these networks, it lacks the temporal resolution to capture their millisecond-scale coordination. b) Temporal perspective: a schematic ERP waveform illustrating the N400 (semantic processing,��~400 ms) and P600 (syntactic/integrative processing,��~600 ms) components that reflect the rapid timecourse of language comprehension. c) Topographical maps illustrating the spatial distribution of ERP effects, reproduced from Schacht et��al. (2014, Fig 1B and 1C): the main effect of semantics in the N400 window (top) and the grammaticality-violation effect in the P600 window (bottom). Licensed under CC BY; cropped from the original. While EEG provides precise temporal information about language processing stages, it offers limited spatial resolution for localizing these processes within the distributed brain networks shown in (a). This methodological divide has hindered our understanding of how spatially distributed language networks achieve their rapid, coordinated function during real-world comprehension.

]]> /valiant/2026/01/28/harmonizing-10000-connectomes-site-invariant-representation-learning-for-multi-site-analysis-of-network-connectivity-and-cognitive-impairment/ Wed, 28 Jan 2026 17:12:54 +0000 /valiant/?p=5704 Newlin, Nancy R.; Kim, Michael E.; Kanakaraj, Praitayini; McMaster, Elyssa M.; Cho, Chloe; Gao, Chenyu; Hohman, Timothy J.; Beason-Held, Lori L.; Resnick, Susan M.; O’Bryant, Sid E.; Phillips, Nicole R.; Barber, Robert Clinton; Bennett, David Alan; Barnes, Lisa Laverne; Biber, Sarah A.; Johnson, Sterling C.; Archer, Derek B.; Li, Zhiyuan; Zuo, Lianrui; Moyer, Daniel C.; & Landman, Bennett Allan. (2025).��.��Journal of Medical Imaging,��12(6), 64001.��

Diffusion magnetic resonance imaging data collected across different studies often vary because of differences in scanners, software, and acquisition protocols, which can introduce unwanted technical effects that interfere with biological analysis. This is especially challenging in large multi site studies of Alzheimer’s disease, where both technical variation and true disease related changes are present. In this study, the authors developed a harmonization approach that learns low dimensional representations of brain structural connectivity that are insensitive to imaging site, scanner type, and acquisition settings, while still preserving information related to cognitive status. They used a conditional variational autoencoder, a type of deep learning model that compresses data into a latent space while controlling which information is retained or removed. The model was trained on diffusion imaging data from 6956 individuals across 9 cohorts and 38 imaging sites, totaling nearly 12,000 scans, including participants with normal cognition, mild cognitive impairment, and Alzheimer’s disease. The learned representations successfully removed statistically significant site related effects across 12 brain network connectivity measures and improved the accuracy of predicting cognitive diagnosis from 68 percent to 73 percent. The method showed consistent performance across multiple data configurations, demonstrating that representation learning can effectively reduce scanner related confounding while strengthening biologically meaningful signals in large neuroimaging studies.

Fig.��1

A total of 11,927 connectomes are pooled together across 9 cohorts and 35 unique imaging acquisitions (for a total of 38 “sites”). (a)��If we naively combine all data and create��t-SNE plots, there are obvious clusters driven by site. We compare cognitive diagnosis differences in one network connectivity measure, brain modularity, computed from these uncorrected connectomes. (b)��Site-wise differences drive the detected differences and result in false findings. (c)��In comparison, the��t-SNE features of our learned representation have decreased site information, and the (d)��resulting modularity values are harmonized and provide clear, significant (p<0.001) trends with cognitive diagnosis.

The first annual International Society of Tractography (IST) debate, held in Corsica in 2024, focused on key challenges and controversies in tractography, a technique that maps brain connections using diffusion MRI. The debate centered on the provocative statement: “Tractography cannot give us anything we can’t get from an atlas template.” On one side, white matter atlas templates were highlighted as standardized, population-based maps of the brain that are especially useful for studying brain structure and making group comparisons. On the other side, subject-specific tractography was presented as a powerful approach that reconstructs the unique brain connections of each individual, allowing an in vivo “virtual dissection” of white matter pathways. This article introduces these concepts, summarizes the arguments for and against the statement, and, while acknowledging the strengths of both methods, underscores the unique value of tractography in advancing our understanding of individual brain connectivity.

Fig. 1

Schematic illustration of distinctions between atlas vs. subject-specific tractography. Atlas: volumetric or streamline-based atlases can be warped to an individual subject in a process called label/streamline propagation, enabling the identification and labeling of white matter pathways or regions. Subject-specific Tractography: bundle segmentation can be performed using a targeted approach (often using inclusion/exclusion regions of interest), or by generating a whole brain set of streamlines and subsequent processing to dissect and segment the streamlines of interest. We note that bundle dissection can be performed using manual regions of interest, or even making use of atlases to compare shape/location similarity when extracting and labelling subject-specific streamlines

Expert programmers tend to be better at coding, but it’s still unclear exactly why. Some researchers have used brain scans (like fMRI) to study how programmers think while doing specific coding tasks, such as understanding code. However, those studies haven’t found consistent brain differences based on experience. One possible reason is that focusing only on tasks may limit the brain areas that get activated during the scans.

In neuroscience, another approach is to study the brain while it’s at rest—that is, when a person is just lying still in the scanner. This “resting-state” brain activity reflects how the brain is naturally organized and can reveal long-term effects of experience. In this study, researchers analyzed resting brain scans from 150 people, including 96 programmers, to see how programming experience might shape brain networks.

They found that programmers showed stronger connections between brain regions related to language, math, and attention over time. In contrast, non-programmers showed more connections in areas linked to social and emotional thinking. The study also found that with more years of programming experience, there was less connection between two specific brain areas involved in reading visuals and speaking.

These findings suggest that the brain may reorganize itself with programming experience, particularly in ways that support logic, language, and focus.

Fig. 1:��

Group-level functional connectivity measures for (a) programmers and (b) non-programmers. Values in these matrices represent the functional connectivity (i.e., Pearson correlation) from every cortical brain region to every other cortical brain region. Correlation coefficients in these figures were Fisher’s z-transformed to normalize their distribution. These values now denote scaled correlation values following a normal distribution between -1 and 1.

Most brain imaging studies have focused on gray matter—the part of the brain that processes information—while paying less attention to white matter, which connects different brain regions. This study looks at how gray and white matter work together by introducing two new measurements. The first, called gray-white matter functional connectivity, tracks how closely the activity in gray and white matter is timed together. The second, called the gray-white blood oxygenation-level dependent (BOLD) power ratio, compares how strong the brain signals are in gray matter versus white matter.

The study found that gray-white matter functional connectivity follows patterns related to how the brain is wired and organized, especially in areas that control movement and basic senses—suggesting that signals move efficiently between gray and white matter. On the other hand, the power ratio showed the opposite pattern, with higher values in brain regions involved in more complex thinking, which may mean that those areas require more energy. The power ratio also increased with age, from 8 to 21 years old, suggesting that as the brain develops, its energy needs shift.

Together, these two measurements show how gray and white matter work together differently—one focused on signal clarity, the other on energy use—helping us better understand brain development and communication.

Fig. 1: Overview of functional contrast across the gray-white matter boundary.

a��The white matter surfaces are shown from a lateral view, highlighting the primary motor cortex in blue.��b��Mesh surface of mid-thickness cortex (red) and gray-white matter boundary (blue).��c��Schematic of functional contrast measures normal to the surface. For a given vertex v_ion the boundary surface, its corresponding GM vertex (G) and WM point (W) are identified along the line perpendicular to the surface.��d��Schematic of functional connectivity measures tangential to the surface. ReHo is calculated for each vertex v_i��on the mid-thickness surface by measuring the correlation of BOLD signals among neighboring vertices, which are connected along the tangential direction to the surface. ReHo regional homogeneity.

Short association fibers (SAFs) are small nerve fibers located in the outer layers of the brain’s white matter that help nearby regions of the brain communicate with each other. Although they play a key role in local brain connections, SAFs have not been widely studied until recently, due to the lack of imaging methods able to capture this part of the brain in detail. Understanding how SAFs develop, especially in comparison to longer white matter tracts that connect distant brain regions, is important for learning how the brain matures from childhood through young adulthood.

This study set out to do three things: map where SAFs are located in relation to long-range white matter tracts, describe how these fibers change during normal brain development, and explore how changes in SAFs relate to changes in the long-range connections. Researchers studied brain scans from 616 participants, ages 5.6 to 21.9 years old, using advanced imaging methods called diffusion tensor imaging (DTI) and neurite orientation dispersion and density imaging (NODDI). These techniques allow scientists to look at how brain tissue is structured and how water moves through it, which provides insights into how brain connections grow and change.

The results showed that both SAFs and long-range white matter tracts followed similar patterns of development. Some features, such as the way water diffuses in the brain (measured by MD, AD, and RD), decreased with age, while other features (like FA, ICVF, ISOVF, and ODI) increased, which is consistent with healthy brain development. However, SAFs and long-range tracts also showed important differences in their development, suggesting that the outer and deeper parts of the brain’s white matter mature in slightly different ways.

In addition to age-related findings, the study also found differences between males and females in several brain measures. These findings highlight that brain development can vary by sex. Overall, this study offers new insights into how the brain’s short- and long-range connections grow during childhood and adolescence, and provides a strong foundation for future research into unusual brain development or disease.

Figure 1

Age and sex distribution in the study cohort of 616 participants (279 M, 337 F), ranging from 5.6 to 21.9 years old with a mean age of 14.5 years.

Alzheimer’s disease (AD) is characterized by the buildup of β-amyloid (Aβ), tau, and neurodegeneration, which negatively impact cognitive function by disrupting brain networks. While the effects of cortical Aβ and tau on brain connectivity are known, it is unclear whether plasma biomarkers of AD pathology are related to these changes. In this study, we explored whether plasma biomarkers related to AD (Aβ42/40, phosphorylated tau [pTau-181]), astrogliosis (glial fibrillary acidic protein [GFAP]), and neuronal injury (neurofilament light chain [NfL]) are associated with changes in brain network connectivity over time, and whether these changes are linked to cognitive decline.������

We measured plasma biomarkers using advanced assays and collected functional brain network data through resting-state fMRI from cognitively healthy adults in the Baltimore Longitudinal Study of Aging. We also assessed cognitive performance at each scan. Using statistical models, we investigated how these biomarkers related to changes in connectivity, whether these relationships varied by amyloid status, and whether changes in connectivity were connected to cognitive decline.������

Our findings (n = 486; average age = 65.5 years) revealed that higher levels of GFAP at baseline were linked to faster declines in connectivity in several brain networks, including somatomotor, limbic, and frontoparietal networks. We also found that the relationship between certain biomarkers, like NfL, and network connectivity changes was stronger in amyloid-positive individuals. Additionally, among a subset of participants with multiple scans, changes in connectivity were associated with cognitive decline, although these results did not hold after correcting for multiple comparisons.������

This study suggests that plasma biomarkers of amyloidosis, astrogliosis, and neuronal injury are associated with declines in brain network connectivity, particularly in amyloid-positive individuals. However, the study had limitations, including the absence of certain tau biomarkers (pTau-217 and pTau-231) and comparative PET data.����

��

Figure 1 Study Time Line and Study Exclusion Flowchart��

(A) BLSA 3T resting-state fMRI scans were initiated in 2011. Baseline visits were considered the first visit for which participants had concurrent plasma biomarker measurement and 3T fMRI. From the baseline biomarker measurement, amyloid status (low vs high) was defined using an ROC analysis of Aβ42/40 for optimal prediction of amyloid (PiB)-PET positivity in BLSA participants (n = 212). Cognitive assessments began before 2011, but only those with concurrent resting-state fMRI scans were included. Finally, using a predefined cortical parcellation, 7 intrinsic functional networks were derived for each resting-state fMRI scan. (B) Study flow diagram for participant selection. Aβ = β-amyloid; BLSA = Baltimore Longitudinal Study of Aging; CVA = cerebral vascular accident; GFAP = glial fibrillary acidic protein; NfL = neurofilament light chain; PiB = 11C-Pittsburgh compound-B; PD = Parkinson disease; pTau-181 = tau phosphorylated at threonine-181; ROC = receiver operating characteristic; sMRI = structural MRI. Created in BioRender. Dark H (2024) .��

Alzheimer’s disease (AD) is characterized by the buildup of certain proteins in the brain and damage to brain cells, which can lead to memory and thinking problems. This study aimed to explore whether blood tests measuring these proteins could predict changes in brain network activity over time in people who do not yet show signs of cognitive decline. Researchers looked at a group of participants from the Baltimore Longitudinal Study of Aging, measuring levels of proteins related to AD (amyloid-β, tau), brain cell injury, and other factors like astrogliosis (a type of brain cell response to damage). They then compared these blood markers to brain scans to see how changes in brain connectivity were related to these biomarkers over several years.��

The study followed 490 participants (average age 65) over an average of 4 years. They found that higher levels of certain proteins (Aβ42/40, GFAP, and NfL) were linked to faster changes in brain connectivity in several brain networks, particularly in people with higher amyloid levels. These changes in brain function were more pronounced in those with abnormal amyloid buildup in the brain. In contrast, no significant changes were seen in people without amyloid buildup. Additionally, the study found that brain network activity at the beginning of the study could predict later changes in cognitive abilities, such as memory and verbal skills.��

In conclusion, for cognitively healthy individuals, certain blood markers can predict future changes in brain network activity, especially in those with higher amyloid levels. These changes in brain function could potentially contribute to future cognitive decline.��

Arachnoid granulations (AGs) are small structures that help manage cerebrospinal fluid (CSF) flow. In Parkinson’s disease (PD), issues with CSF flow might relate to changes in AG size and function. This study compared AG size and number between healthy individuals and those with PD and examined their links to symptoms. Results showed that people with PD had larger and more AGs than healthy controls. These AG changes were linked to motor and cognitive symptoms and to objective sleep problems, though not to self-reported sleep issues. This suggests AGs may play a role in PD-related changes in the brain.��

Fig. 1: Group differences in arachnoid granulation (AG) total volume (mm3). A��Significantly increased total AG volume in baseline Parkinson disease scans compared to healthy controls.��B��Significantly increased AG number in Parkinson disease compared to healthy controls.��CSignificantly increased maximum AG volume in Parkinson disease compared to healthy controls.��DSignificantly increased mean AG volume in Parkinson disease compared to healthy controls. Violin plots are shown with conventional boxplots and individual data points overlaid. Corrected��p-values are shown.��

This study explored how the brain’s connections between white and gray matter (WM-GM) are altered in early (preclinical) Alzheimer’s disease (AD) and how these changes relate to amyloid buildup and thinking abilities. Researchers analyzed brain connectivity and network properties in 344 participants, including those with preclinical AD, AD dementia, and healthy controls.��

They found that preclinical AD is linked to weaker connections in specific WM-GM regions and less separation of brain networks related to control, attention, and movement. Many of these changes were tied to higher levels of amyloid beta, a hallmark of AD. One specific connection showed a direct link to cognitive decline. In AD dementia, these disruptions were more widespread and strongly associated with amyloid levels and cognitive issues. These findings highlight how changes in WM-GM connectivity offer insights into brain dysfunction early in the AD process.��

FIGURE 1 ����

Atlas-defined WM bundles and GM parcels, illustration of bipartite graph model, split of biadjacency matrix into functional
networks. A, Forty-six atlas-defined WM bundles (Table S2 in supporting information). B, Two hundred atlas-defined GM parcels are classified into 17 functional networks (Table S3 in supporting information). C, A simplified bipartite model and the projection from weighted bipartite graph to weighted unipartite graph. The biadjacency matrix is the adjacency matrix of a WM–GM FC graph. D, An example illustrating the algorithm of bipartite-to-unipartite projection. E, Split of biadjacency matrix into 17 functional networks and subsequent projections. ACR, anterior corona radiata; ALIC, anterior limb of internal capsule; BCC, body of corpus callosum; CGG, cingulum in the cingulate gyrus; CGH, cingulum hippocampus; CP, cerebral peduncle; CST, corticospinal tract; EC, external capsule; FC, functional connectivity; FX, fornix; FXC, fornix cres; GCC, genu of corpus callosum; GM, gray matter; ICBP, inferior cerebellar peduncle; MCBP, middle cerebellar peduncle; ML, medial lemniscus; PCR, posterior corona
radiata; PLIC, posterior limb of internal capsule; PTR, posterior thalamic radiation; RLIC, retrolenticular limb of internal capsule; SCBP, superior cerebellar peduncle; SCC, splenium of corpus callosum; SCR, superior corona radiata; SFO, superior fronto-occipital fasciculus; SLF, superior longitudinal fasciculus; SS, sagittal stratum; UF, uncinate fasciculus; WM, white matter��