Summary: By combining neuroimaging data with deep learning methods, researchers have created a brain model that reproduces patterns of neurological impairment more realistically than previous approaches.
Source: University at Buffalo
“I call it my ‘chocolate and peanut butter moment.’”
A University at Buffalo neuroimaging researcher has developed a computational model that more closely mirrors real-world patterns of brain impairment than existing simulations. This innovation merges two established analytic approaches to build a digital testing ground that could aid stroke survivors and patients with other brain injuries by helping researchers generate and evaluate hypotheses about localized neurological damage.
“This model is anchored to the brain’s functional connectivity and reproduces realistic patterns of cognitive impairment,” says Christopher McNorgan, assistant professor of psychology in UB’s College of Arts and Sciences. “Because the model reflects how brain regions are connected, we can manipulate it to gain insight into which areas might be damaged in a patient.”
McNorgan is careful to note that this work does not claim to be a complete digital replica of the human brain. “However, the model’s behavior aligns with how the brain performs, suggesting it captures properties that move us toward a biologically plausible facsimile,” he says.
The study offers a robust way to identify and understand brain networks and their functions, potentially opening new avenues for diagnosis and rehabilitation research. Full details of the model and its evaluation are published in the journal NeuroImage.
At the heart of McNorgan’s approach are two complementary techniques: functional connectivity and multivariate pattern analysis (MVPA).
Traditional neuroimaging models typically use general linear methods that evaluate activity at individual brain locations in response to stimuli. Functional connectivity studies, often using fMRI, map how regions are linked by measuring correlated activity; linear models assume direct relationships, for example between a visual stimulus and activation in a visual area.
Linear methods are powerful for identifying which regions respond to specific conditions, but they can miss complex, distributed relationships that emerge across multiple regions. That is where MVPA and other non-linear, machine-learning approaches excel: rather than treating regions independently, MVPA assesses patterns of activity across many areas simultaneously.
MVPA can detect non-linear combinations of features. McNorgan gives a simple analogy: neurons that represent the meaning of a stop sign don’t respond solely to redness or to an octagonal shape; they respond when features coincide. “Non-linear methods like MVPA are central to deep learning systems used in advanced computer vision, such as those behind self-driving cars,” he explains.
Yet MVPA on its own is opportunistic and can conflate coincidence with meaningful correlation. Conversely, functional connectivity provides biologically motivated constraints but may lack sensitivity to distributed, non-linear patterns. Each method has strengths and limitations, and integrating their outputs has traditionally required extensive interpretation by experts.
McNorgan’s key advance is integrating functional connectivity with MVPA within a single, mutually constrained model. By doing so, the model respects real-world connectivity while learning distributed activity patterns—effectively allowing the two approaches to inform and restrain one another.
“It really was my chocolate-and-peanut-butter moment,” McNorgan says. His background in both neuroimaging and computational modeling made the integration apparent in hindsight.
To train and test the models, McNorgan used fMRI data from participants who imagined examples from three categories—tools, musical instruments and fruits—while in the scanner. Eleven subjects visualized and internally heard familiar items (for example, hammers, guitars and apples), producing BOLD activity patterns tied to each category.
“There are distinctive patterns of activity across the brain when people think about one category versus another,” McNorgan says. “We can think of these as neural fingerprints.”
These BOLD patterns were digitized to train multilayer neural networks to recognize the activity signatures associated with each category. After training, the models were tested on unseen activity patterns: significantly above-chance classification indicated the networks had learned generalizable mappings between brain activity and conceptual categories.
To evaluate biological plausibility, McNorgan introduced simulated “virtual lesions” by disrupting activity in regions known to contribute to specific categories. Models that encoded regional functional connections produced classification errors that matched the lesion site: disabling a region implicated in tool representation selectively impaired tool classification while leaving other categories relatively intact. Models trained without the combined constraints did not show such targeted deficits.
“This method reveals how regions that may seem unimportant in isolation can be essential when considered as part of a broader network,” McNorgan says. “Identifying those regions helps explain why a stroke or injury to a particular area can produce selective impairments in distinguishing certain object types.”
Source:
University at Buffalo
Media contact:
Bert Gambini – University at Buffalo
Image credit:
Image is in the public domain.
Original research (open access):
“Integrating functional connectivity and MVPA through a multiple constraint network analysis,” Christopher McNorgan et al., NeuroImage. DOI: 10.1016/j.neuroimage.2019.116412.
Abstract (summary)
The study introduces a novel analysis that unites general linear model-based brain mapping, multivariate pattern analysis, and connectivity approaches by allowing MVPA and functional connectivity to mutually constrain a single model. Using a multisensory imagery task, multilayer neural networks learned category representations from cortical BOLD patterns, with some models explicitly encoding regional connectivity. Models that included functional connections achieved higher classification accuracy and produced lesion-specific, category-appropriate impairments. Replication in an independent dataset supports the approach. The authors conclude that mutually constrained network analyses yield more parsimonious, biologically plausible models that can facilitate discovery in cognitive neuroscience.