2. La criminología crítica como paradigma
2.5 El enfoque materialista como nuevo paradigma normativo
2.5.2 El problema de cómo conocer el punto de vista de las clases
The extraction of neuron ROIs and their corresponding calcium traces from raw imaging data is not trivial. Prior to any data processing, the raw imaging movies must first be corrected for any motion artifacts to ensure that putative calcium transients are not the result of shifts in the imaging plane. Motion corrected movies can then be analyzed by a number of different cell extraction algorithms to identify putative neuron ROIs and their corresponding calcium traces. While a plethora of method exist for two- photon imaging data, here I will limit my discussion to three different algorithms used for cell detection in single-photon miniscope data and their advantages/disadvantages:
Principal Components Analysis-Independent Components Analysis (PCA-ICA), Constrained Non-Negative Matrix Factorization for microendoscope data (CNMF-E),
and A Technique for Extracting Neuronal Activity from Single Photon Neuronal Image Sequences (TENASPIS).
PCA-ICA was the initial method utilized by Ziv et al. (2013) in their seminal work demonstrating the feasibility of using miniscopes for long-term recordings of neuronal activity. PCA-ICA uses advanced mathematical techniques to identify clusters of imaging pixels that reliably explain the variance observed in pixel intensities (putative neurons). Its main advantage is that it is a principled algorithm which has been reliably used to detect and confirm a number of hippocampal phenomena first discovered with electrophysiology, e.g. place fields (Rubin et al., 2015; Ziv et al., 2013). However, PCA- ICA was originally employed using GCaMP3 with a relatively sparse neuronal
population relative to more current GCaMP variants. As such, it is not ideally suited for disambiguating highly overlapping neurons that occur in hippocampal imaging (Hamel et al., 2015). In fact, one report using simulated data found that as the correlation between neuron ROIs increased, reflecting more overlap between ROIs and more similar ROI shapes, PCA-ICA eventually became incapable of separating neuron ROIs (Zhou et al., 2018). Additionally, it also ended up producing much poorer reconstructions of the actual calcium trace (Zhou et al., 2018). The end result is that PCA-ICA may end up detecting many fewer cells and including more crosstalk than other algorithms (like CNMF-E and TENASPIS) that use image segmentation techniques to identify neurons (Hamel et al., 2015). Additionally, the details of the cell selection process are effectively hidden from the experimenter, making troubleshooting of imaging artifacts and poor quality traces difficult to impossible (Pnevmatikakis et al., 2016). Thus, while a good tool for analyzing
early miniscope imaging data, the shortcomings of PCA-ICA highlighted the need for more sophisticated analysis tools.
CNMF-E was adapted from a similar algorithm designed for two-photon day analysis partially to address this issue. Briefly, CNMF-E assumes that the fluorescence changes observed in each imaging movie are composed of four basic components: 1) spatial footprints for each neuron ROI each with a 2) corresponding time series of calcium activity restrained to that ROI, 3) background composed of a homogeneous global fluorescence changes encompassing the entire FOV and heterogeneous local fluorescence changes related to the activity of neighboring neurons and those
above/below the viewing plane, and 4) temporally and spatially uncorrelated noise (Zhou et al., 2018). These four components are then assembled into an equation whose output represents the observed intensity in each pixel across the entire movie. Placing reasonable constraints on each of these components (e.g. ROIs approximating the size/shape of neurons and a sparsely active, non-negative time series of activity within each ROI, etc.) makes this equation solvable. CNMF-E has a significant number of advantages over PCA-ICA. First, it is open-source and thus accessible for troubleshooting and quality control. Second, it produces higher SNR traces due to accurate subtraction of background fluorescence and noise. Third, it produces more accurate results when compared to PCA- ICA on simulated data, missing less neurons and providing much more accurate ROI and calcium trace estimates, particularly as the SNR of the imaging movie decreases. Most importantly, it significantly outperformed PCA-ICA in accurately disambiguating highly overlapping neurons and faithfully reproducing their actual spatial footprints and calcium
traces. Most of the disadvantages are related to the implementation of the algorithm. For example, there are a number of parameters that must be carefully calibrated for proper implementation of CNMF-E, and these can vary drastically for different brain regions, acquisition rates, pixel sizes, etc. Additionally, utilizing CNMF-E on large/long movies can overwork the resources of all but the most top-end commercially available
computers. However, these disadvantages are tractable problems to fix. Thus, CNMF-E improved significantly upon PCA-ICA for the extraction of neuronal signal from miniscope imaging data.
Like CNMF-E, TENASPIS originated to address the inability of PCA-ICA to disambiguate signal from highly overlapping neurons. The full implementation of
TENASPIS is discussed in depth below (see 2.4.3.1). Briefly, TENASPIS first employs a spatial filtering technique to assist in disambiguating neighboring neurons. It then uses image segmentation techniques to identify highly correlated activity patterns in each imaging frame, corresponding to activity of putative neurons, and then to connect these patterns across frames to create neuron ROIs and their corresponding traces. TENASPIS is capable of accurately extracting neuronal data from imaging movies, as demonstrated by replications of reliable place field activity (section 2) and coding of elapsed time (Mau et al., 2018) in hippocampal neurons. Additionally, TENASPIS can accurately
disambiguate calcium events from neurons with overlapping neuron ROIs (see Figure 2.1 and Figure 3.1). Like CNMF-E, it is open source and employs a heuristic that facilitates quality control. However, unlike CNMF-E it has not undergone rigorous benchmarking against simulated data and other cell extraction methods. Additionally, since it was
developed explicitly for extracting data from imaging the mouse hippocampus, TENASPIS requires adjustment of many parameters to accommodate data extraction from other brain regions. Overall, TENASPIS is one viable method for extracting neuronal data from miniscope imaging movies, and is used to analyze data throughout this thesis, but it requires cross-validation with other methods to remain a
viable/competitive in the future.
Regardless of the method used to extract putative ROIs and their events, matching calcium activity to spiking remains difficult. In some two-photon recordings, it is
possible deduce single action potentials from calcium activity (Chen et al., 2013). However, the accuracy of spike detection via the use of deconvolution algorithms declines with the signal-to-noise ratio (Pnevmatikakis et al., 2016), making it difficult to infer individual spikes from imaging data obtained via single-photon imaging.