The Automatic Neuroscientist: A framework for optimizing experimental design with closed-loop real-time fMRI

Romy Lorenz, Ricardo Pio Monti, Inês R. Violante, Christoforos Anagnostopoulos, Aldo A. Faisal, Giovanni Montana, Robert Leech

TL;DR

This paper introduces the "Automatic Neuroscientist," a novel framework that uses real-time fMRI and optimization algorithms to automatically design experiments that evoke specific brain states. This approach dramatically speeds up fMRI data acquisition and allows for a more comprehensive understanding of the complex relationships between stimuli and neural responses.

What Problem Does This Paper Address?

Traditional functional neuroimaging (fMRI) typically focuses on mapping how a single task or a few tasks activate specific brain regions (one-to-many mapping). However, the brain is far more complex: the same neural systems can be activated by many different tasks, and a single task can engage multiple brain regions. This leads to a "many-to-many" mapping challenge. Current methods struggle to systematically explore this, making it difficult to understand the full functional role of brain regions and how different cognitive processes interact. This is akin to trying to understand a complex system by only looking at a few of its inputs and outputs.

Methods

The core of the "Automatic Neuroscientist" is a closed-loop system that uses real-time fMRI to dynamically adjust experimental parameters.

  • Real-time fMRI: Brain images are acquired and pre-processed on the fly.
  • Optimization Algorithms:
    • Study 1 (SPSA): A Simultaneous Perturbation Stochastic Approximations (SPSA) algorithm was used. This method is suitable for optimizing functions that are not analytically defined and are corrupted by noise, which is common in fMRI. It iteratively proposes new stimulus combinations based on evaluating the brain's response.
  • Study 2 (Bayesian Optimization): A more sophisticated Bayesian optimization approach was employed. This method builds a probabilistic model of the brain's response to stimuli, allowing for more efficient exploration of the parameter space and better handling of noisy data. It uses a Gaussian process prior and an "expected improvement" acquisition function to guide the search for optimal stimuli.
  • Stimulus Generation: Parametric variations of audio-visual stimuli were created. For visual stimuli, this included altering frame rate, image size, saturation, spatial blurring, and adding Gaussian noise. For auditory stimuli, variations involved noise-vocoded speech with different numbers of frequency bands and added Gaussian noise.

Datasets

Not explicitly discussed as pre-existing datasets for training. The study utilizes data collected from human participants during real-time fMRI scanning. The stimuli themselves were parametrically generated based on established methods for visual and auditory stimuli.

Key Insights and Ideas

  • Flipping the fMRI Paradigm: The paper fundamentally shifts the focus from "what brain regions does this task activate?" to "what stimuli/tasks evoke a desired brain state?"
  • Efficiency Gains: Real-time adaptation dramatically reduces scanning time compared to exhaustive search methods.
  • Many-to-Many Mapping: The framework is designed to tackle the complex many-to-many relationship between stimuli and brain activity, offering a more holistic view of brain function.
  • Individualized Experiments: The approach can be tailored to individual subjects, potentially capturing unique brain responses and optimizing for personalized interventions.
  • Handling Noise: Both SPSA and Bayesian optimization methods are designed to be robust to the inherent noise in fMRI data.

Key Results

  • Study 1 (SPSA): Achieved convergence to the hypothesized optimal audio-visual stimulus combination in 11 out of 14 runs in under 10 minutes. This is significantly faster than the estimated half-hour needed for an exhaustive search of the 100 possible states.
  • Study 2 (Bayesian Optimization): Demonstrated rapid and accurate estimation of the entire experimental parameter space. Using just the first run (6.3 minutes), it achieved a statistically significant and accurate mapping of the hypothesized optimum, which would have taken over two hours for an exhaustive scan.
  • Model Performance: In real-time, the Bayesian optimization model showed decreasing uncertainty as more observations were collected, indicating it was learning effectively.
  • Inter-individual Differences: The study highlighted variations in how subjects' brains responded and how the algorithm estimated the parameter space, suggesting the potential for personalized experimental designs.

Limitations and Future Work

  • SPSA Limitations: The SPSA algorithm primarily aims for convergence and doesn't provide a full description of the underlying parameter space. It can also be susceptible to noisy outliers.
  • Bayesian Optimization: While more advanced, the efficiency of mapping the parameter space depends on its complexity and the signal-to-noise ratio of the target brain state. For subtle distinctions, more time might be needed.
  • Future Work:
    • Developing online stopping criteria for Bayesian optimization to further enhance efficiency.
  • Extending the "Automatic Neuroscientist" to any target brain state, including functional connectivity networks.
  • Combining real-time fMRI with other techniques like dynamic functional connectivity estimation.
  • Exploring applications in higher-level cognitive tasks, clinical rehabilitation (neurofeedback, therapy tailoring), brain stimulation optimization, and assessing individual preferences.
  • Investigating the optimal number of runs needed for robust parameter space estimation.

Summary updated

Abstract

Highlights • Mapping between cognitive task and neural processes is often not straightforward. • Current fMRI approach is not sufficient for holistic investigation of this mapping. • We propose a novel framework combining real-time fMRI and optimization techniques. • We validate our closed-loop method and find it to be highly efficient and accurate. • This framework is generalizable to numerous innovative applications. Abstract Functional neuroimaging typically explores how a particular task activates a set of brain regions. Importantly though, the same neural system can be activated by inherently different tasks. To date, there is no approach available that systematically explores whether and how distinct tasks probe the same neural system. Here, we propose and validate an alternative framework, the Automatic Neuroscientist , which turns the standard fMRI approach on its head. We use real-time fMRI in combination with modern machine-learning techniques to automatically design the optimal experiment to evoke a desired target brain state. In this work, we present two proof-of-principle studies involving perceptual stimuli. In both studies optimization algorithms of varying complexity were employed; the first involved a stochastic approximation method while the second incorporated a more sophisticated Bayesian optimization technique. In the first study, we achieved convergence for the hypothesized optimum in 11 out of 14 runs in less than 10 min. Results of the second study showed how our closed-loop framework accurately and with high efficiency estimated the underlying relationship between stimuli and neural responses for each subject in one to two runs: with each run lasting 6.3 min. Moreover, we demonstrate that using only the first run produced a reliable solution at a group-level. Supporting simulation analyses provided evidence on the robustness of the Bayesian optimization approach for scenarios with low contrast-to-noise ratio. This framework is generalizable to numerous applications, ranging from optimizing stimuli in neuroimaging pilot studies to tailoring clinical rehabilitation therapy to patients and can be used with multiple imaging modalities in humans and animals. Previous article in issue Next article in issue
BibTeX
@article{lorenz2016automatic,
  author = {Romy Lorenz and Ricardo Pio Monti and Inês R. Violante and Christoforos Anagnostopoulos and Aldo A. Faisal and Giovanni Montana and Robert Leech},
  title = {{The Automatic Neuroscientist: A framework for optimizing experimental design with closed-loop real-time fMRI}},
  journal = {NeuroImage},
  year = {2016},
  doi = {10.1016/j.neuroimage.2016.01.032},
}
Figure