Last updated on May 25th, 2018. For full publication list, see google scholar. Links to our software are available here & github.
In-vivo delineation of human brain networks and areas
Information processing occurs via the transformation of neural signals across brain networks. Resting-fMRI is a powerful tool allowing the non-invasive, simultaneous, interrogation of multiple brain networks in living individuals. We applied unsupervised machine learning to resting-fMRI from 1000 participants, generating canonical parcellations of the cerebral cortex, cerebellum and striatum into distributed large-scale networks (Yeo2011; Buckner2011; Choi2012). Our parcellations are widely used as references to study human brain organization and disorders. We have extended our network-level parcellations to areal-level parcellations, which comprises hundreds of regions approximating classical cortical areas with distinct function, connectivity, architectonics and topography (Schaefer2018). Our parcellations were more homogeneous (better) than four widely used parcellations in multiple task-fMRI and resting-fMRI datasets across diverse acquisition and preprocessing protocols. Moving beyond population-level parcellations, we developed an approach to delineate individual-specific cortical networks (Kong2018). Using just one fMRI session (10 minutes), our approach generated individual-specific parcellations of comparable quality to two state-of-the-art approaches using five sessions (50 minutes), which might be potentially useful for patients unable to tolerate longer scan time.
Information processing occurs via the transformation of neural signals across brain networks. Resting-fMRI is a powerful tool allowing the non-invasive, simultaneous, interrogation of multiple brain networks in living individuals. We applied unsupervised machine learning to resting-fMRI from 1000 participants, generating canonical parcellations of the cerebral cortex, cerebellum and striatum into distributed large-scale networks (Yeo2011; Buckner2011; Choi2012). Our parcellations are widely used as references to study human brain organization and disorders. We have extended our network-level parcellations to areal-level parcellations, which comprises hundreds of regions approximating classical cortical areas with distinct function, connectivity, architectonics and topography (Schaefer2018). Our parcellations were more homogeneous (better) than four widely used parcellations in multiple task-fMRI and resting-fMRI datasets across diverse acquisition and preprocessing protocols. Moving beyond population-level parcellations, we developed an approach to delineate individual-specific cortical networks (Kong2018). Using just one fMRI session (10 minutes), our approach generated individual-specific parcellations of comparable quality to two state-of-the-art approaches using five sessions (50 minutes), which might be potentially useful for patients unable to tolerate longer scan time.
- The organization of the human cerebral cortex revealed by intrinsic functional connectivity. Yeo BTT*, Krienen FM*, Sepulcre J, Sabuncu MR, Lashkari L, Hollinshead M, Roffman JL, Smoller JW, Zöllei L, Polimeni JM, Fischl B, Liu H, Buckner RL. Journal of Neurophysiology, 106(3):1125–1165, 2011 [pdf]
- The organization of the human cerebellum revealed by intrinsic functional connectivity. Buckner RL, Krienen FM, Castellanos A, Diaz JC, Yeo BTT. Journal of Neurophysiology, 106:2322-2345, 2011 [pdf]
- The organization of the human striatum revealed by intrinsic functional connectivity. Choi EY, Yeo BTT, Buckner RL. Journal of Neurophysiology, 108(8):2242-2263, 2012 [pdf]
- Opportunities and limitations of functional connectivity MRI. Buckner RL, Krienen FM, Yeo BTT. Nature Neuroscience, 16:832-837, 2013 [pdf]
- Local-Global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Schaefer AL, Kong Ru, Gordon EM, Laumann TO, Zuo XN, Holmes AL, Eickhoff SB, Yeo BTT. Cerebral Cortex, 29:3095-3114, 2018 [pdf]
- Imaging-based parcellations of the human brain. Eickhoff SB, Yeo BTT, Genon S. Nature Reviews Neuroscience, 19:672-686, 2018 [pdf]
- Spatial topography of individual-specific cortical networks predicts human cognition, personality and emotion. Kong R, Li J, Sun N, Sabuncu MR, Schaefer A, Zuo XN, Holmes A, Eickhoff SB, Yeo BTT. Cerebral Cortex, in press [pdf]
Functional architecture of the human brain
Comprehensive delineation of brain networks (previous contribution) provides the foundations for new insights into human brain organization. Using resting-fMRI (from 1000 participants) and task-activation coordinates (from 10,449 experiments), we showed that sensory-motor regions were functionally specialized, i.e., they participated in one resting-state network (Yeo2014) or cognitive process (Yeo2015). By contrast, the association cortex was characterized by complex zones ranging from highly specialized to highly flexible. Functionally flexible regions participated in multiple cognitive processes to different degrees (Yeo2015). Each cognitive process was supported by a constellation of strongly connected brain regions, suggesting that the brain consists of autonomous modules, each executing a discrete cognitive function. Complex behavioral tasks yielded greater activity at connector nodes (hubs with diverse connectivity across modules), while activity at local nodes (within-module hubs) remained constant. Thus, hubs with diverse inter-module connectivity are necessary for the functioning of modular biological networks (Bertolero2015). Furthermore, we discovered a mechanistic model in which connector hubs tune the connectivity of their neighbors to be more modular, while allowing for task appropriate information integration across communities, thus increasing global modularity and cognitive performance (Bertolero2018).
Comprehensive delineation of brain networks (previous contribution) provides the foundations for new insights into human brain organization. Using resting-fMRI (from 1000 participants) and task-activation coordinates (from 10,449 experiments), we showed that sensory-motor regions were functionally specialized, i.e., they participated in one resting-state network (Yeo2014) or cognitive process (Yeo2015). By contrast, the association cortex was characterized by complex zones ranging from highly specialized to highly flexible. Functionally flexible regions participated in multiple cognitive processes to different degrees (Yeo2015). Each cognitive process was supported by a constellation of strongly connected brain regions, suggesting that the brain consists of autonomous modules, each executing a discrete cognitive function. Complex behavioral tasks yielded greater activity at connector nodes (hubs with diverse connectivity across modules), while activity at local nodes (within-module hubs) remained constant. Thus, hubs with diverse inter-module connectivity are necessary for the functioning of modular biological networks (Bertolero2015). Furthermore, we discovered a mechanistic model in which connector hubs tune the connectivity of their neighbors to be more modular, while allowing for task appropriate information integration across communities, thus increasing global modularity and cognitive performance (Bertolero2018).
- The organization of local and distant functional connectivity in the human brain. Sepulcre J, Liu H, Talukdar T, Martinocorena I, Yeo BTT, Buckner RL. PLoS Computational Biology, 6(6): e1000808. doi:10.1371, 2010 [pdf]
- Estimates of segregation and overlap of functional connectivity networks in the human cerebral cortex. Yeo BTT, Krienen FM, Chee MWL, Buckner RL. Neuroimage, 88:212-227, 2014 [pdf]
- Reconfigurable state-dependent functional coupling modes cluster around a core functional architecture. Krienen FM, Yeo BTT, Buckner RL. Philosophical Transactions of the Royal Society B, 369:20130526, 2014 [pdf]
- Borders, map clusters, and supra-areal organization of the visual cortex. Buckner RL, Yeo BTT. Neuroimage, 93:293-297, 2014 [pdf]
- Functional specialization and flexibility in human association cortex. Yeo BTT, Krienen FM, Eickhoff SB, Yaakub SN, Fox PT, Buckner RL, Asplund CL, Chee MWL. Cerebral Cortex, 25:3654-3672, 2015 [pdf]
- The modular and integrative functional architecture of the human brain. Bertolero MA, Yeo BTT, D'Esposito M. Proceedings of the National Academy of Sciences USA, 112:E6798-E6807, 2015 [pdf]
- Data-driven extraction of a nested model of human brain function. Bolt T, Nomi JS, Yeo BTT, Uddin LQ. Journal of Neuroscience, 37:7263–7277, 2017 [pdf]
- The diverse club. Bertolero MA, Yeo BTT, D'Esposito M. Nature Communications 8:1277, 2017 [pdf]
- Topographic organization of the cerebral cortex and brain cartography. Eickhoff SB, Constable RT, Yeo BTT. NeuroImage, 170:332–347, 2018 [pdf]
- A mechanistic model of connector hubs, modularity and cognition. Bertolero MA, Yeo BTT, Bassett DS, D'Esposito M. Nature Human Behavior, 112: E6798, 2018 [pdf]
Individual differences in behavior and disorder
Machine learning is critical for precision medicine. For example, current mental disorder categories are based on symptom checklists and not carving nature by its joints. Subtypes within disorders and overlaps across disorders suggest the existence of shared neurobiological factors across disorders. We have utilized unsupervised machine learning to discover latent factors in late onset Alzheimer’s Disease (AD) dementia. The factors were associated with distinct atrophy patterns, as well as distinct memory and executive function decline trajectories among dementia patients and at-risk nondemented participants (Zhang2016). There is significant ongoing work in the lab on discovering factors (or subtypes) underlying heterogeneity in neurological and psychiatric disorders. While unsupervised machine learning is useful for discovering new insights into mental disorders and individual differences, supervised machine learning is useful when we know what we want to predict, e.g., behavioral traits, disease progression and treatment responses. For example, we have successfully utilized the resting-fMRI of well-rested individuals to predict whether they will be vulnerable to the deleterious effects of sleep deprivation(Yeo2015). We have also utilized deep recurrent neural networks to predict AD progression up to five years into the future (Nguyen2018).
Machine learning is critical for precision medicine. For example, current mental disorder categories are based on symptom checklists and not carving nature by its joints. Subtypes within disorders and overlaps across disorders suggest the existence of shared neurobiological factors across disorders. We have utilized unsupervised machine learning to discover latent factors in late onset Alzheimer’s Disease (AD) dementia. The factors were associated with distinct atrophy patterns, as well as distinct memory and executive function decline trajectories among dementia patients and at-risk nondemented participants (Zhang2016). There is significant ongoing work in the lab on discovering factors (or subtypes) underlying heterogeneity in neurological and psychiatric disorders. While unsupervised machine learning is useful for discovering new insights into mental disorders and individual differences, supervised machine learning is useful when we know what we want to predict, e.g., behavioral traits, disease progression and treatment responses. For example, we have successfully utilized the resting-fMRI of well-rested individuals to predict whether they will be vulnerable to the deleterious effects of sleep deprivation(Yeo2015). We have also utilized deep recurrent neural networks to predict AD progression up to five years into the future (Nguyen2018).
- Disruption of cortical association networks in schizophrenia and psychotic bipolar disorder. Baker JT, Holmes AJ, Masters GA, Yeo BTT, Krienen FM, Buckner RL, Öngür D. JAMA Psychiatry, 71:109-118, 2014 [pdf]
- Functional connectivity during rested wakefulness predicts vulnerability to sleep deprivation. Yeo BTT, Tandi J, Chee MWL. Neuroimage 111:147-158, 2015 [pdf]
- Bayesian model reveals latent atrophy factors with dissociable cognitive trajectories in Alzheimer’s disease. Zhang XM, Mormino EC, Sun N, Sperling RA, Sabuncu MR, Yeo BTT. Proceedings of the National Academy of Sciences USA, 113:E6535–E6544, 2016 [free download]
- Inference in the age of big data: future perspectives on neuroscience. Bzdok D, Yeo BTT. NeuroImage, 155:549-564, 2017 [pdf]
- The human cortex possesses a reconfigurable dynamic network architecture that is disrupted in psychosis. Reinen JM, Chen O, Hutchison OY, Yeo BTT, Anderson KM, Sabuncu MR, Öngür D, Roffman JL, Smoller JW, Baker JT, Holmes AJ. Nature Communications 9:1157, 2018 [pdf]
- Modeling Alzheimer’s disease progression using recurrent neural networks. Nguyen M, Sun N, Alexander D, Feng J, Yeo BTT. International Workshop on Pattern Recognition in Neuroimaging, in press.
- Is deep learning better than kernel regression for functional connectivity prediction of behavior? He T, Kong R, Holmes AJ, Sabuncu MR, Eickhoff SB, Bzdok D, Feng J, Yeo BTT. International Workshop on Pattern Recognition in Neuroimaging, in press.
Multi-scale neuroscience
The human brain is a complex system spanning from micrometer (cellular) to centimeter (fMRI) scales. We have contributed to bridging the gap between microscale and macroscale brain organizations. For example, we found that genes enriched in the supragranular layers of the human cerebral cortex (relative to mouse) were expressed in a topography reflecting broad cortical classes (sensory/motor, paralimbic, associational) and associated network properties (Krienen2016). Therefore, molecular innovations of upper cortical layers may be important for the evolution of long-range corticocortical projections. In addition to genetics, we have been exploiting machine learning algorithms to invert large-scale biophysical models. The estimated biophysical model parameters in turn provide insights into the large-scale and cellular organization of the human brain (Wang2018).
The human brain is a complex system spanning from micrometer (cellular) to centimeter (fMRI) scales. We have contributed to bridging the gap between microscale and macroscale brain organizations. For example, we found that genes enriched in the supragranular layers of the human cerebral cortex (relative to mouse) were expressed in a topography reflecting broad cortical classes (sensory/motor, paralimbic, associational) and associated network properties (Krienen2016). Therefore, molecular innovations of upper cortical layers may be important for the evolution of long-range corticocortical projections. In addition to genetics, we have been exploiting machine learning algorithms to invert large-scale biophysical models. The estimated biophysical model parameters in turn provide insights into the large-scale and cellular organization of the human brain (Wang2018).
- Transcriptional profiles of supragranular-enriched genes associate with corticocortical network architecture in the human brain. Krienen FM, Yeo BTT, Ge T, Buckner RL, Sherwood C. Proceedings of the National Academy of Sciences USA, 113:E469-E478, 2016 [pdf]
- A spotlight on bridging microscale and macroscale human brain architecture. van den Heuvel M, Yeo BTT. Neuron, 93:1248-1251, 2017 [pdf]
- Gene expression links functional networks across cortex and striatum. Anderson KM, Krienen FM, Choi EY, Reinen JM, Yeo BTT, Holmes AJ. Nature Communications 9:1428, 2018 [pdf]
- Inversion of a large-scale computational model reveals a cortical hierarchy in the dynamic resting human brain. Wang P, Kong R, Liegeois R, Deco G, van den Heuvel MP, Yeo BTT. Science Advances, 5:eaat7854, 2019 [free download]
Image processing and statistics
We have contributed towards resolving brain imaging processing and statistical issues. For example, an in-depth analysis of statistical issues involved in studying fMRI dynamics revealed several surprising results, e.g., first order autoregressive models explain fMRI dynamics better than nonlinear biophysical models (Liegeois2017). As another example, the results of most neuroimaging studies are reported in volumetric or surface coordinate systems. Accurate mappings between the two coordinate systems can facilitate many applications, but there is surprisingly little research on this topic. We found that an improved implementation of an old algorithm (from Yeo2011) for mapping between volumetric and surface coordinate systems worked very well (Wu2018). While the previous examples are specific to neuroimaging, we have also developed general methods applicable beyond neuroimaging. For example, the cerebral cortex is often represented as a 2D sphere, motivating our interest in spherical image processing and registration. We extended a sampling theorem in Euclidean space to the 2-Sphere and utilized the theorem to construct overcomplete wavelets for spherical image processing (Yeo2008). We have also exploited differential geometric techniques and Lie group of diffeomorphisms to develop fast algorithms for image registration (Yeo2009; Yeo2010).
We have contributed towards resolving brain imaging processing and statistical issues. For example, an in-depth analysis of statistical issues involved in studying fMRI dynamics revealed several surprising results, e.g., first order autoregressive models explain fMRI dynamics better than nonlinear biophysical models (Liegeois2017). As another example, the results of most neuroimaging studies are reported in volumetric or surface coordinate systems. Accurate mappings between the two coordinate systems can facilitate many applications, but there is surprisingly little research on this topic. We found that an improved implementation of an old algorithm (from Yeo2011) for mapping between volumetric and surface coordinate systems worked very well (Wu2018). While the previous examples are specific to neuroimaging, we have also developed general methods applicable beyond neuroimaging. For example, the cerebral cortex is often represented as a 2D sphere, motivating our interest in spherical image processing and registration. We extended a sampling theorem in Euclidean space to the 2-Sphere and utilized the theorem to construct overcomplete wavelets for spherical image processing (Yeo2008). We have also exploited differential geometric techniques and Lie group of diffeomorphisms to develop fast algorithms for image registration (Yeo2009; Yeo2010).
- On the construction of invertible filter banks on the 2-sphere. Yeo BTT, Ou W, Golland P. IEEE Transactions on Image Processing, 17(3):283--300, 2008 [pdf]
- Effects of registration regularization and atlas sharpness on segmentation accuracy. Yeo BTT*, Sabuncu MR*, Desikan R, Fischl B, Golland P. Medical Image Analysis, 12(5):603--615, 2008 [pdf]
- DT-REFinD: diffusion tensor registration with exact finite-strain differential. Yeo BTT, Vercauteren T, Fillard P, Peyrat J-M, Pennec X, Golland P, Ayache N, Clatz O. IEEE Transactions on Medical Imaging, 28(12):1914--1928, 2009 [pdf]
- Spherical demons: fast diffeomorphic landmark-free surface registration. Yeo BTT*, Sabuncu MR*, Vercauteren T, Ayache N, Fischl B, Golland P. IEEE Transactions on Medical Imaging, 29(3):650--668, 2010 [pdf]
- Learning task-optimal registration cost functions for localizing cytoarchitecture and function in the cerebral cortex. Yeo BTT, Sabuncu MR, Vercauteren T, Holt D, Amunts K, Zilles K, Golland P, Fischl B. IEEE Transactions on Medical Imaging, 29(7):1424--1441, 2010 [pdf]
- A generative model for image segmentation based on label fusion. Sabuncu MR*, Yeo BTT*, Van Leemput K, Fischl B, Golland P. IEEE Transactions on Medical Imaging, 29(10):1714--1729, 2010 [pdf]
- Interpreting temporal fluctuations in resting-state functional connectivity MRI. Liegeois R, Laumann TO, Snyder AZ, Zhou HJ, Yeo BTT. Neuroimage 163:437–455, 2017 [pdf]
- Accurate nonlinear mapping between MNI volumetric and FreeSurfer surface coordinate systems. Wu J, Ngo GH, Greve DN, Li J, He T, Fischl B, Eickhoff SB, Yeo BTT. Human Brain Mapping, in press [pdf]