Data-driven segmentation of cortical calcium dynamics
Source: Weiser, Sydney C., et al. “Data-driven segmentation of cortical calcium dynamics.” PLOS Computational Biology 19.5 (2023): e1011085.
Full Article: Data-driven segmentation of cortical calcium dynamics | PLOS Computational Biology
Overview
In this study, researchers demonstrate the use of independent component analysis to separate neural activity from artifacts in widefield recordings of neuronal cortical calcium dynamics. They develop a random forest classifier that automatically distinguishes between neural signals and artifacts with high accuracy. By analyzing the spatial and temporal metrics of the components, they create a functional segmentation map of the mouse cortex, revealing distinct domains that represent underlying neural signals.
Background
Optical techniques have been used to study neuronal activity in various organisms, including the mammalian visual cortex. Calcium imaging allows for monitoring neural activity across the cortical surface; however, it is challenging to separate neural signals from other confounding sources. Wide-field cortical calcium imaging provides detailed dynamics but is affected by artifacts like movement and blood flow changes. This study proposes an independent component analysis (ICA)-based approach to identify and filter out artifacts, enabling the extraction of neural signals. The ICA algorithm separates signal sources based on their spatial and temporal properties, leading to the identification of distinct neural components. The study also explores the impact of recording parameters on signal quality and demonstrates the creation of data-driven maps for functional segmentation. These techniques enhance the filtering, segmentation, and time series extraction of wide-field calcium imaging videos, facilitating the study of complex cortical dynamics.
Results
A method for recording neural activity patterns in the cortex of awake behaving adult mice using a genetically encoded calcium indicator called GCaMP6s. The indicator is expressed in all neurons under the control of the Snap25 promoter. The process involves transcranial recording of fluorescence from the mouse’s cortex using blue wavelength light and capturing the emitted green light with a high-resolution camera. The recorded video is then processed to analyze the neural activity patterns.
The data obtained from the video is subjected to a spatial Independent Component Analysis (ICA) decomposition, which separates the recorded signals into different components. These components can be classified into three major categories: neural components, artifact components, and noise components. Neural components represent specific areas of cortical tissue and exhibit distinct spatial and temporal properties. Artifact components can arise from various sources such as blood vessels, movement, and optical distortions. Noise components do not have a specific spatial domain and lack temporal structure.
The ICA decomposition allows for the reconstruction of the video data using different combinations of the identified components. For example, a filtered video can be constructed by excluding artifact components. The artifact components can also be reconstructed separately to verify that the desired neural signal was not removed during the filtering process.
The quality of the ICA decomposition is influenced by the spatial and temporal resolution of the recorded data. Higher spatial resolution improves the separation of neural and artifact signals, while lower resolution compromises the separation. On the other hand, temporal resolution has less impact on the separation quality. The duration of the recorded video also affects the segmentation of the cortex into distinct components, with 20 minutes being a sufficient duration to capture most of the neural and artifact components.
To classify the extracted components as neural or artifact, manual scoring is performed based on visual inspection of their spatial and temporal properties. Neural components typically exhibit globular spatial representations and dynamic temporal activations. Vascular artifact components can be identified by their vascular-like spatial patterns, while other artifact components may show diffuse spatial representations with smaller or sparse temporal activations.
The researchers used a random forest classifier and trained it on a subset of the data, achieving high accuracy, precision, and recall metrics. They also projected the features onto a two-component space and observed distinct clusters for neural signals and artifacts. The classifier exhibited reliable confidence in its decisions.
To assess the efficacy of the classifier, the researchers tested it on novel data that was not part of the training set. The results of the classifier on the novel data were similar to the subset classifier, indicating consistent performance. However, there were some misclassifications, primarily occurring at the edges of the region of interest or within specific brain regions.
The text also discusses the importance of filtering the global mean signal to account for artifacts before reintroducing it during data reconstruction. The researchers found that applying a high-pass filter with a 0.5 Hz cutoff minimized global slow oscillations associated with artifacts.
Additionally, the researchers used the spatial characteristics of the independent components (ICs) to create domain maps of the cortical surface. These domain maps represented the spatial distribution of signals within the cortex and were used to extract time courses from the cortical surface. The time courses extracted from the domain maps outperformed those generated from other downsampling methods in terms of representing the total signal variation in the filtered data.
Furthermore, the researchers investigated the relationships between ICs, domains, and the full-resolution filtered data. They found high correlations between ICs and their corresponding domains, indicating the domain maps captured the underlying neural signal. They also quantified the independence of each domain from its surrounding neighborhood and found that higher-order domains tended to be more independent.
Finally, the text mentions that animal-specific domain maps can be regionalized based on reference maps and domain features, suggesting that the domain shapes reflect underlying functional units in the cortex.
Overall, the study demonstrates the effectiveness of machine learning algorithms in classifying neural signals and artifacts and highlights the utility of domain maps for extracting time courses and analyzing neural activity.
Discussion
Wide-field calcium imaging is a technique that uses genetically encoded calcium indicators to visualize neuronal activity. However, isolating neural signal sources from various artifacts in the data has been a challenge. The authors propose an ICA-based processing pipeline that effectively removes artifacts and allows for the segmentation of cortical patches based on their functional units.
The authors compare their approach to previous methods used in the field, such as point correlation studies and spike-triggered average investigations. They also discuss the impact of recording parameters, such as temporal and spatial down sampling, on the quality of the results. They find that temporal down sampling has little effect, but spatial down sampling can lead to increased artifacts and less informative components.
The passage also addresses the influence of non-neuronal signals, particularly hemodynamics, on the population dynamics observed in wide-field calcium imaging. The authors discuss the use of hardware-based solutions and computational methods to minimize these artifacts. They emphasize the importance of removing artifacts from the data to improve data quality and decrease variability.
The authors describe the process of segmenting the cortex using ICA and discuss the stability of the resulting independent components (ICs) over time. They note that the number and types of activations greatly influence the resulting set of ICs. They also highlight the challenges of interpreting the results, such as the inability to separate layer-specific signals and the presence of residual hemodynamic influence.
The passage concludes with the proposal of creating domain maps based on the reconstructed functional data obtained from the ICs. These maps represent patches of cortex with the average reconstruction of functional neural data. The authors discuss the potential applications of these maps, including more computationally demanding spatial models and time series analysis.
FREE ChatGPT Document Q&A: Get questions answered about any document type of any length!
Plug: Please purchase my book ONLY if you have the means to do so, I usually do not advertise, but I am struggling to stay afloat. Imagination Unleashed: Canvas and Color, Visions from the Artificial: Compendium of Digital Art Volume 1 (Artificial Intelligence Draws Art) — Kindle edition by P, Shaxib, A, Bixjesh. Arts & Photography Kindle eBooks @ Amazon.com.
