(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Data-driven segmentation of cortical calcium dynamics [1]

['Sydney C. Weiser', 'Department Of Molecular', 'Cell', 'Developmental Biology', 'University Of California Santa Cruz', 'Santa Cruz', 'California', 'United States Of America', 'Brian R. Mullen', 'Desiderio Ascencio']

Date: 2023-05

Demixing signals in transcranial videos of neuronal calcium flux across the cerebral hemispheres is a key step before mapping features of cortical organization. Here we demonstrate that independent component analysis can optimally recover neural signal content in widefield recordings of neuronal cortical calcium dynamics captured at a minimum sampling rate of 1.5×10 6 pixels per one-hundred millisecond frame for seventeen minutes with a magnification ratio of 1:1. We show that a set of spatial and temporal metrics obtained from the components can be used to build a random forest classifier, which separates neural activity and artifact components automatically at human performance. Using this data, we establish functional segmentation of the mouse cortex to provide a map of ~115 domains per hemisphere, in which extracted time courses maximally represent the underlying signal in each recording. Domain maps revealed substantial regional motifs, with higher order cortical regions presenting large, eccentric domains compared with smaller, more circular ones in primary sensory areas. This workflow of data-driven video decomposition and machine classification of signal sources can greatly enhance high quality mapping of complex cerebral dynamics.

Researchers have been able to record from a large population of neurons across the cortex using calcium indicators in awake, behaving mice; however, many confounding neuronal signals (ie. neurons from different depths) or tissue dynamics (ie. blood flow) influence these recordings. Our custom pipeline utilizes algorithms to identify distinct signals and spatially segment the signal sources into components that can then be characterized. From these components, we show that neuronal signals are distinct from non-neuronal artifacts; further, we are able to remove the artifacts to clean the neuronal signal. The remaining components are spatial segments of the neuronal signals; we use them to create a data-driven map of functional units, which we call domains. We characterize these domains between subsequent recordings and show that they are highly similar within the same animal, given a long enough recording. This data-driven map can provide information about the limitations one can make from the specific recordings. Furthermore, it will enhance our ability to understand the effects on functional maps from animals that do not have a reference map, including mapping functional changes in development, differences between genetically mutated or varying strains of mice.

Funding: This work was supported by Startup funds from University of California, Santa Cruz, Division of Physical and Biological Sciences, grants from the National Institutes of Health, USA (NIH T32 GM 133391) to S.C.W. and (NIH T32 GM 864620) B.R.M, and by a Hellman Fellows Fund Award to J.B.A. Funding for D.A. was provided by the UCSC Maximizing Access to Research Careers (MARC) program (T32-GM007910) and the UCSC Initiative for Maximizing Student Development (IMSD) (R25-GM058903). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Copyright: © 2023 Weiser et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

We also explore the resolution-dependent effect of signal extraction on ICA quality, and find a quantified increase in ICA signal separation for collecting wide-field calcium imaging at mesoscale resolution. Additionally, using neural components, we generate data-driven maps that are specific to functional borders from individual animals. We use these maps to extract time series from functional regions of the cortex, and show that this method for time series extraction produces a reduced set of time series while optimally representing the underlying signal and variation from the original dataset. Together, these methods provide a set of optimized techniques for enhanced filtering, segmentation, and time series extraction for wide-field calcium imaging videos.

Here we present an ICA-based workflow that isolates and filters artifacts from calcium imaging videos, with principled exploration of each component to identify each signal source necessary to reduce the contamination resulting from these physiological dynamics. ICA is a nonparametric unsupervised machine learning (ML) technique that can identify each signal source in densely sampled (5.5 million pixels per frame) calcium imaging videos based on their spatially co-activating pixels and temporal properties. The global mean time course was initially subtracted and stored, thereby allowing ICA to decompose each signal distinct from global effects. The decomposition results in hundreds of neural source components per hemisphere that are distinctly de-mixed from artifact source signals. Our concurrent analysis of control wide-field imaging data corroborates the identification of artifact signal sources and gives insight into the structure of neuronal calcium dynamics across neocortex.

ICA is a blind source separation algorithm utilized on multichannel data to reveal underlying signal sources. In utilizing this algorithm, we first make the assumption that confounding artifacts and neural calcium signals behave with differing temporal and spatial properties; therefore, artifacts can be removed to result in data that is exclusively neural. ICA has had great result de-mixing resting state fMRI data that requires no a-priori information about the signals of interest [ 29 – 31 ]. Secondly, we assume that the calcium signals collected from structured populations of neurons will produce repetitive and consistent network activation, allowing the algorithm to isolate functional activity based on underlying cytoarchitecture. As such, we can apply ICA decomposition to gain understanding of the functional networks across the cortex.

Eigendecompositions can be used to identify and filter components of signal [ 20 – 22 ], present a flexible method of filtering that is not hardware dependent, and be applied to any video dataset regardless of the recording hardware or parameters. An eigendecomposition pipeline that utilizes non-negative matrix factorization has been developed to explore the functional activities across wide-field imaging of the cortex, but this method is limited by the use of a reference map and cannot separate artifact signals from neural activation [ 23 ]. Independent Component Analysis (ICA) [ 24 ] has the potential to overcome these limitations and has been previously applied to fMRI and EEG data with varying success; for example, identifying both intrinsic connectivity networks rather than individual areas and artifacts that represent large-scale effects rather than spatially localized effects [ 25 – 28 ]. We hypothesize that this is due to the lower density of spatial sampling in fMRI and EEG data.

It is common to use sensory stimulation to identify specific regions in the neocortex, then align a reference map based on the location of these defined regions [ 13 , 15 , 16 ]. Even if these maps are reliable for locating primary sensory areas, they often lack specificity for higher order areas, or even completely lack sub-regional divisions. This is especially true in areas with a high degree of interconnectedness and overlapping functionality, such as motor cortex [ 17 ]. Moreover, there is evidence that the shape and location of higher order regions can vary from subject to subject [ 18 , 19 ]. Improper map alignment or misinformed regional boundaries can lead to a loss in dynamic range between signals across a regional border. In order to extract the most information from a recorded dataset, the level of parcellation must reflect the quality and sources present within the data. Collectively, these considerations demonstrate that a flexible data-driven method is necessary; furthermore, it must also respect functional boundaries of the cortex and be sensitive to age, genotype and individual variation.

Recording parameters need to be taken into consideration to ensure high signal quality and understanding the limitations of wide-field calcium dynamics need to be addressed quantitatively. Researchers have recorded wide field calcium dynamics at frame rates ranging from 5-100Hz [ 9 , 11 , 12 ]. In addition, spatial resolution varies between different researchers’ setups, but is typically in the range of 256x256 to 512x512 pixels (0.06 to 0.2 megapixels) for the entire cortical surface, and is often further spatially reduced for processing [ 9 , 11 , 13 ]. Election of resolution is often dependent on the video observer’s perceived quality of the data or available computational resources, rather than a quantified comparison of signal content. Further, considerations need to be taken on the signal source to understand the impact on signal quality [ 14 ].

Wide-field cortical calcium imaging provides a unique combination of spatially and temporally resolved dynamics across the cortical surface, with scale ranging from complex activation patterns in high-order circuits to discrete activations hundreds of micrometers in diameter to whole cortical lobe activity patterns [ 8 , 9 ]. However, these techniques are affected by issues common to all optical imaging recording. Body or facial movements can create large fluctuations in autofluorescence of the brain and blood vessels, which produce significant artifacts in the data. Vascular artifacts are commonly seen due to vasodynamics and the resulting changes in blood flow required to meet the energy demands of surrounding tissue. Fluid exchange between vascular and neural tissue causes cortical hemodynamics, resulting in region specific changes of optical properties among cerebral lobes [ 10 ]. Further, though the skull is fixed to a specific location during the experiment, slight brain movements occur within the cranium, thereby influencing the recordings. Any optical property differences that originate from the experimental preparation may be highlighted in the dataset as signal due to changes in tissue contrast.

Optical techniques have long been used to monitor the functional dynamics in sets of neuronal elements ranging from isolated invertebrate nerve fibers [ 1 , 2 ] to entire regions of mammalian visual cortex in vivo [ 3 – 5 ]. Imaging calcium flux with calcium sensors [ 6 , 7 ] allows for monitoring transcranial neural activity across the cortical surface of the mouse with high enough spatiotemporal resolution to identify sub-areal networks of the neocortex [ 8 , 9 ]. These techniques have the potential to map supracellular group function at unprecedented resolution and scale across the neocortical sheet in awake behaving mice; however, identifying neural signals from calcium imaging sessions is challenging due to numerous confounding signal sources.

Results

To record neural activity patterns in the cortex of awake behaving adult mice, we transcranially recorded fluorescence from a mouse that has the genetically encoded calcium indicator, GCaMP6s, expressed in all neurons under the control of the Snap25 promoter [32]. We expose and illuminate the cranium with blue wavelength light and capture emitted green light with a sCMOS camera at high spatial resolution (2160x2560 pixels, 5.5 megapixels; ∼6.9 μm/pixel). To observe the spatiotemporal properties of the recorded activity patterns, we crop the video to only neural tissue, and compare the change in fluorescence over the mean fluorescence: ΔF/F over time (Fig 1A and 1B). In order to identify components associated with artifacts and hemodynamic responses, similar data was recorded and processed in three sets of age matched control mice: cx3cr1 GFP (microglia; mGFP), adhl1 GFP (astrocyte; aGFP), and the non-transgenic C57black/6 (Bl6) mice.

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. ICA separates calcium data into its underlying signal components. (A) Recording schematic and fluorescence image of transcranial calcium imaging preparation. Full image with masked regions of interest (ROI, dashed lines) shows regions used in ICA decomposition. (B) Sample video montage of raw video frames after ΔF/F calculation. (C) ICA video decomposition workflow. The ΔF/F movie without the mean is decomposed into a series of statistically independent components that are either neural, artifact, or noise associated (not displayed). Each component has an associated time course from the ICA mixing matrix. Neural components can be rebuilt into a filtered movie (rICA). Alternatively, artifact components can be rebuilt into an artifact movie. Circular panels show higher resolution spatial structure in example in the rightmost components. https://doi.org/10.1371/journal.pcbi.1011085.g001

ICA separates signal sources from high resolution data A spatial ICA decomposition on a video, wherein the global mean was subtracted, produces a series of spatial independent components (sIC) and a mixing matrix, which captures the component’s influence at each frame in the video (Fig 1C). The components are sorted by temporal variance and oriented so that they all represent positive spatial effects (See Methods). The independent components can be sorted into 3 major categories based on their spatiotemporal properties: neural components, artifact components, and noise components (not shown). Group analysis on these components was performed and discussed later in this manuscript. Neural components represent a distinct area of cortical tissue, which we refer to as its cortical domain. The spatial morphology of these neural components can vary in both spatial extent and eccentricity. Neural components are also frequently found to contain multiple domains, which have similar enough activation patterns to be identified as a single neural component. In the examples in Fig 1C, the second neural component appears to represent a secondary somatosensory network, with multiple domains on the right hemisphere, and a small mirrored domain on the left hemisphere. Artifact components can take many forms, including those from blood vessels, movement, and optical distortions on the imaging surface. The left two artifact examples (Fig 1C) likely represent hemodynamics from the superior sagittal sinus vein (left, center) and blood flow through the middle cerebral artery (right) [33]. A very high-resolution map of the vessel patterns can be rebuilt from these components, with branching structures as small as 12 μm in diameter (shown in Fig 1C, right). Noise components lack a spatial domain and have little to no temporal structure. Signal and artifact components can be sorted manually in the provided graphical user interface (S1 Fig) or with a machine learning classifier. Video data can be reconstructed using any combination of these components. In particular, a filtered video can be constructed by excluding all artifact components. The artifact movie can also be reconstructed to verify that desired signal was not removed with the artifact filtration (S1 Video).

Spatiotemporal metrics can be derived from each component to assess the classification of each signal source Using the ICA decomposition from 20-minute duration videos, we inspected each set of experimental components from both controls and GCaMP6 expressing mice to classify each as neural or artifact (Fig 4A). Neural components typically have globular spatial representation with highly dynamic properties. Vascular artifact components can be easily visually identified by the vascular-like spatial representation. Other artifact components that are commonly seen in the components are movement or preparation artifacts. These typically have a diffuse spatial representation with smaller or sparse temporal activations. We manually scored each component in the dataset as an artifact (vascular or other) or neural component (Fig 4B). From all the GCaMP experiments, an average of 73.5 ± 5.9% of the components were identified as neural, where the remaining 26.5±6.3% were artifact (vascular: 8.7 ± 2.7%; other: 17.6 ± 7.1%). GCaMP mice had substantially higher numbers of neural components compared to the controls, resulting in four times as many as the GFP mouse lines and six times the number in Bl6 mice (mean number of neural components GCaMP: 235, mGFP:62, aGFP:54, Bl6: 39). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. Class identity cannot be established by any individual extracted feature. (A) Examples of independent components of neural (n) signal, vascular (v) artifacts, and other (o) artifacts. Components are defined by both the sIC and its temporal fluctuations. Circular windows magnify key portions of the sIC. sIC values represented by colormap from blue to red. Temporal representation is in relative intensity (black time course under the sIC), only 1 minute of the full 20 minutes are shown. (B) A comparison of the number of neural signal (GCaMP: dark blue; controls: light blue) and the artifact components (vascular: red; other: orange) with each animal shown (GCaMP components: N = 12 animals, n = 3851; mGFP components: N = 3, n = 484; aGFP components: N = 3, n = 442; WT components: N = 3, n = 229). (C) Examples of binarization of the sIC. Histogram shows the full distribution of sIC values. The dynamic threshold method to generate binarized masks was used to identify the high sIC signal pixels (yellow) against the gaussian background (blue). Windowed spatial representation shows binarization on the key portions of the sIC. (D) Examples of neural and artifact wavelet analysis shown in the power signal-to-noise ratio (PNR) plots. 95% red-noise cutoff was used to create signal to noise ratio (black dashed lines). (E) Histograms of example spatial metrics derived from GCaMP sIC values, (F) morphometrics from the shape of the binarized primary region, (G) temporal metrics derived from relative temporal intensities, (H) frequency metrics derived from the PNR. https://doi.org/10.1371/journal.pcbi.1011085.g004 We extracted spatial and morphological metrics of the neural and artifact components to characterize spatial feature differences (Fig 4C). We can pull general spatial intensity metrics like global minimum and maximum from the spatial sIC of each component. The largest sIC values correspond to the regions that have the most dynamic change from the data. Given that the shapes of the high pixel intensity values are used by humans to identify their classification, we decided on a dynamic thresholding technique to binarize the sIC. When examining the histogram of intensity values of the neural sIC, there is a large population of pixels centered around zero with a single long tail. We identified all pixels that were unique to the long tail by excluding all values that lie within range of the shorter gaussian tail. From these binarized masks, morphometrics of each primary region of the component can then be quantified, such as the axis lengths or eccentricity of the shape. We characterized temporal dynamics of each component by extracting features from the component time series (Fig 4D). The time series analysis allows us to pull out temporal features of each component, such as standard deviation and global maxima/minima of each component contribution. We performed wavelet analysis on these time series to characterize only highly significant frequencies (S4 Fig). We calculated a power signal-to-noise ratio (PNR) with the 95% quantile of red noise defined by the autocorrelation value of each time series. With this ratio, significant frequencies resulted in a value above 1. After extracting these metrics, we then compared the diverse populations of neural and artifact components, separated between vascular and other, for each feature of interest (Fig 4E–4H). A full list of all metrics and their respective definitions is provided, organizing between spatial, morphometric, temporal, and frequency features (S5 Fig). While the data shows general trends, there is not one single metric that alone could predict the classification of artifacts and neural components. Control neural components are not distinct globular regions like those from the GCaMP line (S6 Fig); rather, they had co-activity with vascular units in the center of its domain. This resulted in the thresholded region being more similar to the vascular artifacts seen in GCaMP components. These identified neural components did not seem to follow known neuronal functional or cytoarchitecture, but were primarily a result of localized vascular and hemodynamic influence. However, we were still able to find example components that only had vascular spatial representation without the surrounding tissue activation. Finally, we found similar artifacts of the other category in the control data that are also present in the GCaMP sets of independent components.

GCaMP mice have strong distinct globular domains that cover the entire cortical surface To investigate how well these metrics captured features of each component, we explored the coverage of the cortical surface with regions identified by the dynamic thresholding technique. By plotting all the contours from one experiment, the major footprint of the component shows its representative space within the brain region (Fig 5A). The majority of defined brain regions are represented by the GCaMP component footprints, with varying amounts of overlap associated with different cortical areas. Control data resulted in sparsely mapped footprints across the cortex. Further, the mapped centroid location of all components that had a thresholded region from all GCaMP experiments shows the neural components have high densities in the sensory regions of the brain (Fig 5B). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 5. Spatial thresholding and frequency data reliably produce neural metrics. (A) Individual experiment preparation with corresponding spatial footprints by class of component: GCaMP neural (dark blue), control neural (light blue), vascular (red), other (orange). (B) All model experiments (N = 12) with corresponding centroid location of each of the class of metrics. Histograms show the resulting average distribution of spatial location across the field of view (error bars are standard deviation between experiments). (C) Individual experiment (same as A), where components are sorted by temporal variance. PNR mapped to each component and organized between the classification of components. (D) Main frequencies seen in each component class between each experimental condition. Dotted lines represent the mean dominant frequency within each animal, where the gray around the mean corresponds to the standard deviation of that animal. The color line corresponds to the grand mean between all experiments. (E) Relative position-based variance of the types of components between experiments and transgenic model, shown as the average and standard deviation between experiments. (F) The percent of components that had footprints and frequency data that was above the noise cut-off, separated by component type and experimental condition. https://doi.org/10.1371/journal.pcbi.1011085.g005 Thresholded GCaMP neural components have high densities in the olfactory bulbs and posterolateral portions of the cortex, including visual, auditory, and somatosensory systems. There is less dense localization of centroids along the anteromedial portions of the cortex, including motor and retrosplenial cortices. Further, in both the GCaMP and control mice, we see the majority of artifact components localize along anatomical brain vasculature. The major venous systems, including the rostral rhinal vein, the superior sagittal sinus, and the transverse sinus, all show high densities of artifact centroid locations [33]. The cerebral arteries are less consistent in localizing the primary domain of their respective components. We see that many of the other artifacts align with the sagittal and lambda cranial sutures [34]. We investigated the effects of wavelet analysis on feature generation by sorting each component within an experiment to its temporal standard deviation value. We then ordered each class of components based on variance and displayed a grayscale heatmap of the significant frequencies in each component across experimental conditions (Fig 5C). Taking the average of each of the global wavelet spectrum across each experiment highlighted the prominent frequencies seen in each classification (Fig 5D). Prominent GCaMP frequencies are between 0.3 to 3.5 Hz, where control dynamics are typically seen between 0-1 Hz. Vascular components tend to have the same frequencies as their neural counterparts, while other components typically have faster frequencies (above 3 Hz), most likely due to motion during the recordings. We looked at the overall distribution of the class of components (neural, vascular, other) with respect to their relative variance (Fig 5E). We found significant shifts in distribution in the types of components based on variance. Neural components were found with high variance, however they significantly tapered off nearing the noise floor. We found the highest percentage of vascular components with high variance, followed by constant low probability throughout the relative variance. The other components had the highest probability close to the noise floor, with low frequency throughout the rest of the relative variance position. Among all GCaMP experiments, 92.3 ± 5.6% of the components had a strong spatial activation, resulting in a thresholded region to assess. However, when we looked at the breakdown of the class of components, we found that 99.8 ± 0.2% of all neural components had a thresholded region (Fig 5F). The artifacts had fewer thresholded components, specifically in the other classification; vascular artifacts had 96.6 ± 3.6% and other artifacts had 60.2 ± 16.2% within their respective class of components that had a domain footprint above the spatial threshold. Of note, we found that components with low variance close to the noise threshold had increased probability of not having a domain threshold. Wavelet analysis between all GCaMP experiments revealed 95.8 ± 3.8% of the components had statistically significant frequencies to assess. We found that 98.4 ± 2.5% of all GCaMP neural components had significant frequencies (Fig 5F). Vascular artifacts had 93.0 ± 6.5% and other artifacts had 86.6 ± 9.4% of components with significant wavelet frequencies within their respective class of components. Overall, this indicated that all metrics can be generated for the vast majority of neural signals and vascular artifacts. However, on average, 40% of the other artifacts do not have morphometrics for their components and 13% are lacking significant frequencies. Further, this also shows changes in the relative spatial and temporal variance distributions of each of these components. The neural and vascular components align with known and predictive anatomy and were found primarily in mapped locations of increased variance. The majority other components have less variance, and therefore less contribution to the original dataset; the ones that had a footprint were typically found along cranial sutures. All control data had fewer footprint and frequency metrics (Fig 5A and 5C).

Spatial metrics best separate neural components from artifacts To build a classifier, we identified metrics that distinguish between neural and artifact components. Correlation of component class to with respect to their metric value and their respective t-statistic between class identified which features are most useful to classify each component (Fig 6A, top). For this process, we randomly selected seven animals from our twelve-experiment dataset for determining features for training the classifier. The remaining five were used as the novel dataset for validating the machine learning performance. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 6. Spatial and morphological metrics are most important to classify components. (A) Correlation and t-statistic between artifact and neural components for each feature (N = 7, n = 2190). Spatial (circles), morphological (triangle), temporal (diamond), and frequency (square) metrics plotted. Cut off values that helped in the selection process are dotted lines, rejected values in gray. Closed points are components that meet requirements. Relative importance metric from the Random Forest classifier plotted against each metric by their respective classes. Selected metrics shown in the list within each type of feature, sorted by greatest t-statistics magnitude. (B) The dataset was parsed into ML modeling dataset (N = 7, n = 2190) that was used to establish the machine learning pipeline and a novel dataset (N = 5, n = 1661) of full experiments that will not influence the classifier. Modeling data was stratified 70/30 split based on classification. 1000 iterations of training the machine learning classifier on selected metrics and validating the machine classification with human classifications. (C) Performance of the ML training, using subsets of the ML modeling dataset. 1000 iterations resulted in accuracy, precision and recall boxplots. (D) 1000 iterations of training on the full ML building dataset was performed and the novel dataset was assessed on its performance. (F) SVD projection of metric data with human classification mapping (top) and the confidence of the ML classifier (bottom). (E) Performance of the classifier on each of the novel datasets, animals plotted separately showing distribution of the 1000 different trained classifiers. (G) Approximate location of false negatives and positives from novel datasets. https://doi.org/10.1371/journal.pcbi.1011085.g006 We trained the random forest classifier with all metrics to identify the importance of each feature (Fig 6A, middle). The features with the greatest t-statistic magnitude had the highest importance for proper classification. In particular, spatial and morphological metrics were found to have the highest relative importance for component classification. The final list of 10 feature metrics utilized in the machine learning process are shown (Fig 6A, bottom).

Machine learning performs as well as human classification We utilized the common approach of hiding a portion of the data from the learning algorithm to validate efficiency of machine learning and establish hyperparameters (Fig 6B). Stratified sorting was used to ensure an equal ratio of artifact and neural signal was placed into each subset. We sampled and trained the random forest classifier 1000 times to see the distribution of results and found that it performed well by all metrics assessed: mean accuracy of 97.1%, mean precision of 98.4%, and mean recall of 97.6% (Fig 6C). All other tested algorithms did similarly well (S7A Fig). After establishing the efficacy of the classifier, we set out to assess the full classifier based on all data points from the machine learning dataset. We projected all features onto the first two components of a singular value decomposition (SVD), mapping both the human classification and the mean classifier confidence for 1000 iterations (Fig 6F). As expected, we saw distinct neural and artifact clusters in feature space. Interestingly, the two different types of artifacts also separated into distinct portions of the projected feature space. The confidence of the classifier showed very few components between the extremes, illustrated by the top binned confidence value distributions for each human classification (S7B Fig). We found 71.2 ± 0.2% of components were binned in highly confident values for neural signals (left-most bin), and 22.7±0.2 were binned in highly confident values for artifacts (rightmost bin) (7.0±0.2% for vascular; 15.7±0.1% for other). This indicates that the classifier exhibits reliable confidence in the decision boundaries. To assess the efficacy of this classifier, we then tested 1000 iterations of novel data—completely new experiments that were not involved in training the classifier (Fig 6D). We plotted the resulting 1000 iterations of each experiment separately (Fig 6E). Notably, we found that the overall results were about the same as the subset classifier: mean accuracy of 96.9%, mean precision of 98.0%, and mean recall of 97.6%. From the histogram of classification frequency, we found similar results to the confidence of the classifier (S7C Fig). Among all components, 69.7±2.0 were confidently classified by machine learning as neural signal (left-most bin), where 25.8±1.4% were confidently classified as artifact (rightmost bin; 7.5±0.4% of vascular; 18.2±1.6% of other). The remaining 5% were mis-classified. We investigated the locations of the components that consistently showed false positive or false negative (Fig 6G). The majority of these components were either on the edge of the region of interest for the cortical hemispheres or within the olfactory bulb.

Global mean needs a high-pass filter to account for removed artifacts before re-addition Removing artifact components will ensure that neural signals are the dominant signal after rebuilding with identified neural components; however, during reconstruction of ICA data, re-addition of the global mean must occur. Thus, we examined the influence of removing artifact components on the global mean and how filtration of the global mean should be considered. For example, vascular artifacts associated with the superior sagittal sinus contribute to the global mean and increase the range of signals recorded during periods of motion (S9 Fig). Assessment of the global mean from GFP control experiments showed pronounced signal in these slower frequency oscillations, suggesting the use of a high pass filter. Indeed, we found that application of a high-pass filter with a 0.5 Hz cutoff minimizes these types of global slow oscillations (S8 Fig). This type of filtration should not be applied to each component individually, as there are regional networks reliant on these slower oscillations [15], but this filtration approach does remove all global oscillations, some of which are neural based fluorescence (discussed below). However, removal of these low frequencies from the global mean improved identification of the cortical patch signal sources that contribute to neural activation (S2 Video).

Domain maps optimize time course extraction from underlying data In addition to their applications for filtering, the components also are a rich source of information about spatial distributions of signal within the cortex. Components across the cortex show a wide diversity of spatial characteristics and represent the detection of independent units of signal. We use the spatial domain footprints of each signal component to create a data-driven ‘domain map’ of the cortical surface by taking a maximum projection through each component layer (Fig 7A). When we compare the number of sIC (249±22) and the number of domains (261±23), we see a significant increase in the number of domains from the number of sIC (Fig 7B; paired t-test p<0.001). Due to the competitive nature of max projection and the frequent number of sIC that have multiple domains, we looked further into the map creation. We were able to match up each domain with its most influential sIC. The majority of sIC had great influence on only one domain (89.5±1.50%). A smaller percentage of sIC had influence on two or more domains (7.55±1.24%). Finally, an even smaller subset of sIC did not have any influence on the domain creation (3.00±0.77%). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 7. Domain maps represent the features of the ICs. (A) Schematic of domain map creation. A maximum projection is taken through each blurred neural component to form a domain map. Mean time courses are extracted from rebuilt filtered movies within a defined domain. (B) Left scatter plot compares the number of neuronal ICs and the resulting number of domains and the dashed line is the identity line. The circled dot indicates the examples used in this figure. Right bar plot shows the fraction of ICs that have major contributions to the number of domains. (C) Examples of ICs that contributed two domains, one domain or had no contribution. The location of the maxima (black circle) of each IC was found and a point correlation map of the rebuilt full resolution filtered movie was produced (green and pink maps). Pixel-wise scatter plot (right) shows the relationship between the spatial IC value and the correlation map value. Dotted line is the threshold value described in Fig 4. Red line indicates the median correlation value of the correlation map that resides in the thresholded IC. (D) IC-Domain correlation with domain map created with filtered mean plotted with respect to the size of each domain. Dark blue corresponds to domains where each IC made more than one domain. Bright blue corresponds to those that made only one. (E) Median correlation value based on point-correlation analysis within each domain (top). Difference in correlation between the center domain with all its adjacent neighbors (gray scale value corresponds to the mean of the median difference value found in each immediately adjacent domain). (F) Example time courses of domain neighborhood. Left domain map identifies the location of each neighborhood with each corresponding time series. Example ICs and point correlation data shown in S10 Fig. https://doi.org/10.1371/journal.pcbi.1011085.g007 At full resolution, there are approximately 1.7 million pixels along the surface of the cortex and olfactory bulb–an impractical number of sources for most network analyses, which work best on 10–300 time series [35]. As such, data-driven domain maps are an optimal method for extracting time courses from the cortical surface. Time series were extracted by averaging the filtered movie under each domain. This results in a series of 261±23 time series per video recording, representing a ∼ 6,500-fold compression rate (Fig 7A; right). In this sense, we no longer have an independent component, but rather the full timeseries of the underlying neural signal. We do this to ensure that every aspect of the neural signal is included in the final analysis, especially those that did not influence the domain map. As we have identified the sIC and its most influenced domain, we are able to make direct comparisons between the temporal dynamics of the sIC and domain. Further, to understand the relationship between sIC, domain, and the full resolution filtered dataset, we performed a pixel-wise point correlation analysis across the full resolution filtered data at the location of neural sIC maxima (Fig 7C). These three pieces of data help us make several direct comparisons to understand the sIC, its corresponding domain, and their relation to the filtered data. First, to assess the independence of each sIC, we make a direct pixel-wise comparison of the sIC value and the matching point correlation map. We see that only the most highly correlated values are associated with the sIC high values. Second, we can directly measure how each temporal feature of the sIC correlates to the rebuilt time series of its most influenced domain (Fig 7D). With the filtered mean re-added, we see that the sIC and its corresponding domain are highly correlated (Pearson’s correlation coefficient: 0.93±0.08). When we identify the multiple domains that arise from the same sIC, they tend to be small domains (dark blue). To assess how well the domain map represents the neural signal in the full resolution filtered data, we can take the median of all correlation values residing in the domain-matched point correlation map (Fig 7E; top). Within each domain, the median pixel-wise correlation coefficient to the sIC maximal location is 0.92±0.07. Further, we can assess how independent each domain is compared to its shared bordered neighborhood. To test for independence, we investigate the null hypothesis; that is, the difference between two point correlation maps associated with adjacent domains should be zero. Two such neighborhood time series (Fig 7F), their corresponding ICs, point correlation maps, and the correlation difference between surrounding and the center domain correlation maps are shown (S10 Fig). In each of the difference maps, we see shifting correlational differences even in adjacent sIC/domains. To quantify how distinct each center domain is to its corresponding surrounding domain, we take the mean of all median differences that reside in the surrounding domains (Fig 7E; bottom). On average, the surrounding domains are 0.15±0.02 less correlated than its center domain. However, this is dependent on the cortical region. Note that the light gray areas reside in primary sensory areas, showing they tend to be more similar; conversely, the darker gray higher order domains tend to be more independent to their neighborhood. To test how well the full filtered video was represented in these time series, we rebuilt ‘mosaic movies’, where each domain is represented by its mean extracted signal at any given time point (Fig 8A and S3 Video). By comparing the borders of the large higher order visual activation, one can appreciate that visually, the data appears more distorted in the Voronoi and grid. To numerically compare whether this method of time course extraction was superior to alternate methods, we calculated the residuals between the two movies. We also compared residuals from mosaic movies rebuilt with either grid or randomly generated Voronoi maps. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 8. Time series extracted from domain maps outperform time series generated from other down sampling methods. (A) Example of a mosaic movie frame rebuilt with respect to each down sampling technique. The non-down sampled filtered movie is represented on the left with subsequent down sampling based on domain, grid or Voronoi maps. (B) Percent total signal of the filtered video represented by extracted time courses. Percent of overall video signal captured in domain maps was calculated for each animal (green circle; N = 8), and compared to signal content from a domain map generated from a separate video from the same animal (green triangle). Percent total signal represented by time courses extracted from grid (blue square) or randomly generated (blue diamond) maps were compared as controls. In the right panel, the percent signal relative to the domain map percent signal was summarized in a box plot. (C) Variation between time courses extracted with each map method was then quantified as a sum signal variation for each experiment. In the right panel, the sum signal variation for each comparison map relative to the optimized domain map sum signal variation was summarized in a box plot. https://doi.org/10.1371/journal.pcbi.1011085.g008 The residuals between the mosaic movies and the filtered movies were compared to the total spatial variation in the filtered movie to quantify the amount of total signal represented by the extracted time courses (Fig 8B, left). In nearly every experiment, the optimized domain map performed better than any other time course extraction method, and accounted for 68±1.2% of the total spatial signal in the filtered video (n = 8). Domain maps generated from different videos from the same animal performed nearly as well as the optimized domain maps created from the compared video (Fig 8B, right). These maps performed significantly better (p = 0.01) than the grid maps, and much better than the Voronoi maps (p < 0.001). Saving these extracted time courses and all associated metadata results in a file size of ∼ 100MB, representing an additional ∼ 60-fold compression compared to saving the full ICA compressed dataset. One potential benefit to accounting for the underlying regions of the brain while extracting time courses is reducing the amount of times that an extracted mean signal is diluted by signal from a neighboring region. Properly restricting time series extraction to statistically independent units should enhance the dynamic range between extracted time series. To test whether domain maps extracted time courses better extract the full range of variation in the cortical surface, we compared the total variation between time courses rebuilt under domain maps from the same video, same animal, or control grid and voronoi maps (Fig 8C, left). When normalized to the performance of the optimized domain map, domain maps from the same animal again had similar performance, but grid and voronoi maps performed significantly worse (p < 0.001; Fig 8C, right). There is a ∼ 15% reduction in signal variation in grid or voronoi maps compared to domain map extracted time courses.

Animal specific domain maps can be regionalized based on reference maps and domain features We were impressed with how the domain shape and structure seemed to capture the functionality of each patch of cortex, so we wanted to investigate the interpretations of each domain shape. We assumed that each sIC would capture underlying functional units established by known mesoscale circuits. As such, we were curious about two key aspects that the domain shape captured underlying circuit structure: domains will have regional, non-uniform restrictions similar to established maps, and subsequent recordings would reliably produce similar domain shapes. To investigate the domain shapes across the cortex, we calculated the same morphometrics used in our machine learning to identify neural sIC for each domain. When we looked for regional differences in domain size across the cortex, we saw a trend for large domains to reside in the anterior and medial portions of the cortex (Fig 9A, left). More impressively, when we plotted the major axis of highly eccentric domains (Fig 9A, right; top quantile of each experiment), we saw regional borders similar to the reference map (Fig 9B). These highly eccentric domains reside by defined structures such as the edge of the cortex, as well as the regions between primary cortical areas. It was not surprising that the eccentric domains reside by the physical structures, but it was surprising we were able to get consistency between individual animals in the locations of the intracortical highly eccentric domains. This led us to think there is a physical structure or circuits that influences these domains. Each domain was first manually sorted into cortical regions based on these shape observations. Second, we looked at the domain correlation maps associated with each cortex. We noted that the frontal cortex tended to have a lateral and median correlation pattern (S10 Fig), allowing us to distinguish between the medial vs lateral motor region (Fig 9C and 9D). Domains did not exhibit uniform spatial characteristics across the neocortex, as seen when we compare the regional differences in spatial characteristics, such as area (ANOVA F = 139, p < 0.001), as well as eccentricity (ANOVA F = 60.0, p < 0.001). Further, we see a significant trend in that higher order and motor regions (R, V+, Ss, Mm, Ml) had larger (p>|t| = 0.000) and more eccentric (p>|t| = 0.000) domains than primary sensory areas (V1, A, Sc, Sb, S). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 9. Domain maps are created from ICA components and are unique to each recording, but highly similar among individual animals. (A) Domain size (circle size) and minor axis (green-purple colorbar) plotted at each position across all hemispheres independently (left). Highest eccentric domains (top quartile) plotted showing the location and direction of its major axis (right; n hemispheres = 14, 7 mice) (B) The Allen Brain atlas map [36] is additionally used for anatomical reference. (C) The final manually assigned region, with associated labels. (D) Domain minor axis and eccentricity by region (grey horizontal line denotes top quantile used in A). 7 mice). (E) Example overlay of one domain map on another from the same animal. Individual domain or region overlap is calculated using the Jaccard index (intersect / union). Population analysis of the Jaccard index for domain (F) and region (G) overlap comparisons. Maps are generated from a different recording on the same animal, a littermate, a non-littermate, or a randomly generated voronoi map. Significance is calculated using a two-way ANOVA, followed by post-hoc t-test analysis with Holm-Sidak correction. Retrosplenial: R; V1: Visual, Higher order visual: V+; Auditory: A; Somatosensory Secondary: Ss; Somatosensory Core: Sc; Somatosensory Barrel: Sb; Somatosensory other: S; Motor medial: Mm; Motor lateral: Ml. https://doi.org/10.1371/journal.pcbi.1011085.g009 To test the meaning of these maps, a series of comparisons were performed. Pairs of maps were overlaid on top of each other (Fig 9E–9G), and every domain was compared to its nearest domain in the comparison map. The Jaccard overlap was calculated for each of these domain pairs, and quantified for each pair of map comparisons. For a null hypothesis, randomly generated Voronoi maps were also compared. Maps generated from different recordings from the same animal were found to be highly overlapping, and hence more similar (Fig 9F; p < 0.001). There was no significant difference in comparisons between littermates vs non littermates. Non-littermate map comparisons were significantly more similar to each other than to voronoi maps (p < 0.001). We additionally quantified whether detected regions were similar across map comparisons. We again found that comparisons between maps from the same animal were highly similar (Fig 9G; p < 0.001), no difference was found between littermates and non-littermates, and comparisons between different animals were significantly more similar than a comparison between a region map and a randomly generated voronoi map (p < 0.001). For comparison, domain maps were created for all analysis done in this manuscript (S11 Fig) and can be compared with domain maps from the down sampling and duration experiments (S12 Fig). In summary, regions and domains are similar between recordings either in the same or on different animals, compared to a null map distribution.

[END]
---
[1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011085

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.