Spatially Structured Sparse Morphological Component Separation for voltage-sensitive dye optical imaging

Hugo Raguet, Cyril Monier, L. Foubert, Isabelle Férézou, Yves Frégnac, Gabriel PeyréJournal of Neuroscience Methods 2015 · DOI

TL;DR

This paper introduces SMCS, a novel method for separating neuronal signals from noise and artifacts in voltage-sensitive dye optical imaging (VSDOI). By leveraging structured sparsity and morphological component analysis, SMCS offers a more robust and detailed view of neural activity compared to existing techniques, paving the way for better understanding of complex brain dynamics.

What Problem Does This Paper Address?

Voltage-sensitive dye optical imaging (VSDOI) is a powerful technique for observing neural activity across large populations of neurons with high temporal resolution. However, the raw VSDOI signal is heavily contaminated by various artifacts, including photobleaching, heartbeat and respiration-induced noise, and camera noise. These artifacts obscure the subtle neuronal signals, making it challenging to accurately interpret brain activity, especially for transient events like propagating waves. Existing methods like blank subtraction and generalized linear models (GLM) struggle to effectively disentangle these complex signals from noise, often leading to loss of information or inaccurate representations of neural dynamics. This paper aims to address this by developing a more sophisticated signal separation method that can better handle the spatiotemporal structure of both the desired neuronal signal and the interfering artifacts.

Methods

The core of the proposed method, SMCS (Spatially Structured Sparse Morphological Component Separation), is a variational framework that extends the GLM approach. Instead of relying solely on predefined regressors, SMCS incorporates non-smooth convex regularizations to enforce sparsity and morphological priors on different components of the signal.

Key methodological aspects include:

  • Sparsity-based Decomposition: The signal is modeled as a sum of components (bleaching, periodic artifacts, neuronal activity), each represented by a dictionary of regressors. The method then seeks a sparse representation, meaning only a few regressors are needed to explain each component.
  • Spatiotemporal Structure: SMCS explicitly leverages the spatial and temporal structure of the signals. For periodic artifacts, it uses an $l_{1,2}$-norm by blocks, assuming that similar frequencies affect neighboring pixels similarly. For neuronal activity, it employs sparse temporal wavelets and total variation for sparse spatial gradients, capturing localized and propagating activity patterns.
  • Morphological Component Analysis: The method incorporates morphological priors, such as assuming neuronal activity is spatially smooth or exhibits wave-like patterns, to guide the separation process.
  • Optimization: A generalized forward-backward splitting algorithm is used to solve the resulting non-smooth convex optimization problem efficiently.
  • Parameter Selection: An automatic parameter selection procedure is employed, which is crucial for the method's robustness and applicability.

Datasets

The paper evaluates the SMCS method on both synthetic and biological datasets:

  • Synthetic Data: A random generative model was created to simulate VSDOI data, mimicking recordings of propagating waves in the cat's visual cortex. This allowed for quantitative evaluation of separation performance by comparing retrieved components to ground truth.
  • Biological Data:
    • Cat Visual Cortex (Area 18): Data from experiments involving full-field drifting luminance gratings were used to assess the method's ability to extract orientation selectivity maps.
  • Mouse Somatosensory Cortex (S1): Data from whisker deflection experiments were used to evaluate the method's performance on capturing rapid, propagating neuronal activity.

Key Insights and Ideas

  • The Importance of Spatiotemporal Structure: The paper strongly emphasizes that effectively separating VSDOI signals requires exploiting both spatial and temporal correlations, a limitation of many previous methods.
  • Structured Sparsity for Artifacts: Applying sparsity constraints not just to the signal but also to the structure of artifacts (e.g., spatial blocks for periodic noise) significantly improves separation.
  • Wavelets as Powerful Neuronal Activity Models: The use of sparse wavelet representations proves effective in capturing the transient and localized nature of neuronal activity, including spontaneous events.
  • Distinguishing Neuronal Activity from Artifacts is Crucial: The paper highlights the difficulty in separating neuronal activity from periodic artifacts, especially when artifacts have complex temporal dynamics or when neuronal signals are weak.
  • Hard Constraints Improve Robustness: Incorporating hard constraints (e.g., non-negativity for bleaching, bounded amplitudes for periodic artifacts) enhances the reliability of the separation, particularly in real biological data.

Key Results

  • Synthetic Data: SMCS demonstrated competitive performance against GLM methods in separating neuronal responses, particularly when considering specific time windows. SMCS showed superior performance in retrieving periodic artifacts.
  • Cat Visual Cortex Data: SMCS successfully generated orientation maps that were comparable to those obtained by simpler methods like blank subtraction and GLM, suggesting it preserves essential functional information. Crucially, SMCS showed less bias of stimulus-related signals into non-neuronal components compared to GLM.
  • Mouse Somatosensory Cortex Data: SMCS effectively removed high-frequency noise and captured rapid, propagating neuronal activity in single trials. It also showed promise in separating spontaneous activity from noise, even when it occurred at different times and locations than evoked activity.

Limitations and Future Work

  • Complexity of Periodic Artifacts: The paper acknowledges that the definition of periodic artifacts is somewhat vague, and neuronal activity can exhibit oscillatory behaviors at various frequencies, making their separation challenging.
  • Blank Data Quality: The performance of SMCS is sensitive to the quality of blank acquisitions, as these are used to model artifacts. If blank data contains significant neuronal signals, it can lead to misclassification.
  • Parameter Tuning: While SMCS offers automatic parameter selection, fine-tuning might still be necessary for optimal performance in specific experimental conditions.
  • Future Work: The authors suggest further investigation into the nature of periodic artifacts, exploring more complex cortical functional properties, and testing the method on data from awake animals, where artifacts might be more pronounced. They also highlight the potential for extending the method to other imaging modalities.

Summary updated

BibTeX
@article{raguet2015spatially,
  author = {Hugo Raguet and Cyril Monier and L. Foubert and Isabelle Férézou and Yves Frégnac and Gabriel Peyré},
  title = {{Spatially Structured Sparse Morphological Component Separation for voltage-sensitive dye optical imaging}},
  journal = {Journal of Neuroscience Methods},
  year = {2015},
  doi = {10.1016/j.jneumeth.2015.09.024},
}
Figure