2.4.1.

Animals and husbandry

All procedures involving living animals were performed in accordance with the guidelines and regulations of the U.S. Department of Health and Human Services and approved by the Institutional Animal Care and Use Committee at the University of California, Santa Barbara. Adult C57Bl/6 mice of both sexes ($>8$ weeks; Jackson Labs) were housed under a 12/12 h dark-light reversed cycle with ad libitum access to food and water.

2.4.2.

Surgery

A 4-mm diameter craniotomy was performed over the visual cortex as previously described.⁴⁷ Briefly, mice were premedicated with a sedative, acepromazine ($2\mathrm{mg}/\mathrm{kg}$ body weight, i.p.), after which they were deeply anesthetized using isoflurane (2% to 3% for induction and 1% to 1.5% for surgery). The mouse’s body temperature was monitored and actively maintained using an electronic heat pad regulated via a rectal probe. Carprofen ($25\mathrm{mg}/\mathrm{kg}$ body weight, s.c.) was administered preoperatively, and lidocaine solution containing epinephrine ($5\mathrm{mg}/\mathrm{kg}$ body weight s.c.) was injected locally before and after the scalp excision. The scalp overlaying the right visual cortex was removed, and a custom head-fixing imaging chamber with a 5-mm diameter opening was mounted to the skull with cyanoacrylate-based glue (Oasis Medical) and dental acrylic (Lang Dental) or dental cement (C&B Metabond, Parkell). Mice were mounted on a custom holder via the headplate chamber, which was filled with a physiological saline containing (in mM) 150 NaCl, 2.5 KCl, 10 HEPES, 2 ${\mathrm{CaCl}}_2$, and 1 ${\mathrm{MgCl}}_2$. A craniotomy was performed using carbide and diamond dental burs on a contra-angle handpiece (NSK).

2.4.3.

Intrinsic signal optical imaging and retinotopic maps

In order to functionally map the visual cortex for targeted injection of viral vectors, intrinsic signal optical imaging (ISOI) was performed using a custom microscope and a CCD camera as previously described.^47,48 During ISOI, mice were lightly anesthetized with 0.5% isoflurane augmented by acepromazine ($2\mathrm{mg}/\mathrm{kg}$ body weight, i.p.), and the body temperature was kept at 37°C. A single-drifting white bar on a black background (3 deg thick, moving in elevation or azimuth direction) was used as previously described⁴⁷ to obtain retinotopic maps based on the intrinsic signals. Retinotopic maps were used to locate V1. The pial vasculature map relative to the retinotopic maps was used to guide targeted injections into V1 at one to two sites.

2.4.4.

Viral injections

After craniotomy and ISOI was performed to map V1, adeno-associated viral (AAV) vectors were injected into V1 under continued isoflurane anesthesia as previously described.^19,23,47 Briefly, 1:1 mixture of pENN.AAV.CamKII 0.4.Cre.SV40 [AAV1; Addgene #105558; diluted at 1:20,000 in phosphate-buffered saline (PBS)] and pGP.AAV.syn.FLEX.jGCaMP7b.WPRE (AAV1; Addgene #104493; original concentration at $\sim {10}^{13}\mathrm{vg}/\mathrm{mL}$) or pGP.AAV.syn.FLEX. jGCaMP8m.WPRE (AAV1; Addgene #162378; original concentration at $\sim {10}^{13}\mathrm{vg}/\mathrm{mL}$) viral particles were injected (65 nL per site; 1 to 2 sites per animal) into V1 with a pulled-glass capillary micropipette using a Nanoliter 2010 controlled by a microprocessor, Micro4 (World Precision Instruments) at 15 nL per min. The glass pipette was left in place for 5 min before retracting to avoid the backflushing of the injected solution. The cranial window was then sealed with a glass cranial plug made up of 4-mm and 3 or 3.5-mm circular coverslips (Warner Instruments) stacked in tandem with a UV-curing optical adhesive (NOA61, Norland).

2.4.5.

Visual stimuli for assessing orientation tuning

During two-photon calcium imaging of dendritic orientation tuning responses, the visual stimulus LCD monitor was covered with a cone-shaped shroud with a small opening at the tip that was positioned around the eye of the mouse. This prevented the imaging pathways from getting contaminated with the visual stimulus light. The stimulus extended from $+20\mathrm{deg}$ to $+124\mathrm{deg}$ in azimuth and from $-10\mathrm{deg}$ to $+42\mathrm{deg}$ in elevation. Stimulus frames consisted of $128\times 128\text{pixels}$, which were smoothly interpolated such that one pixel was equivalent to $0.72{\mathrm{deg}}^2$ in visual space. Visual stimuli were generated using MATLAB and the Psychophysics toolbox.⁴⁹ Orientation tuning was assessed using drifting black and white square wave gratings as previously described (8 or 12 different directions at $0.04\text{cycles}/\mathrm{deg}$ and 2 Hz.⁴⁷ Each grating was presented for 2 or 3 s separated by 3 or 1 s of gray screen in between, respectively. Each sweep of gratings was repeated 5 to 10 times.

2.4.6.

Locomotion

The mouse was allowed to explore a 180-cm long virtual linear maze while headfixed under the objective for two-photon imaging. An encoder was used to feed the locomotion of the mouse to update the instant position within the maze. During closed-loop, the locomotion and the translation of the maze were instantly engaged, while during open-loop, the locomotion and the translation of the maze were decorrelated (the locomotion of a previous trial led the translation of the maze). The virtual environment was displayed in front of the mouse using LED monitors. Three LED monitors were concatenated horizontally by 120-deg angles, which covered the mouse’s visual field that was 150 deg in azimuth and 32 deg in elevation. The position in the maze and the instant locomotion speed of the mouse were recorded. Dendritic spine activity relative to an increase in speed was analyzed.

2.4.7.

Two-photon microscopy imaging

Two-photon imaging of ${\mathrm{Ca}}^{2+}$ transients indicated by GCaMP7b or 8m was performed starting 4 to 6 weeks after AAV injection, using a custom-built two-photon microscope used in prior studies.^17,40,47 Our routine imaging sessions were performed using a 16× Nikon objective (N16XLWD-PF, 0.80 NA) with the two-photon laser set to 910 nm at 60 to 100 mW power measured at the objective. Frame scans were acquired using ScanImage⁵⁰ at either 30 frames per second ($512\times 512\text{pixels}$), 58 frames per second ($512\times 256\text{pixels}$), or 110 frames per second ($512\times 128\text{pixels}$).

3.4.1.

Fully automated versus user supervised semiautomated ROI detection

Accurate measurement of local fluorescent transients of a given biological compartment heavily relies on successful segmentation of ROIs. To this end, we have developed a flexible tool for dendritic spine detection that allows the users to (1) automatically identify all available puncta in the image stack, (2) manually select their preferred ROIs supported by a “point-and-click” style of ROI seed identification, or (3) a combination of the two [Figs. 3(a) and 3(b)]. For all these approaches, AUTOTUNE uses algorithm by which dendritic spine features are segmented based on the correlation map of the intensity transients among neighboring pixels of the seed pixel [Fig. 3(a)]. Users can manually select the seed pixels and/or use the auto spine detection function. The latter identifies seed pixels by a particle detection algorithm based on Gaussian filtering and background intensity estimation.⁵³ Neighboring pixels for generating correlation map are defined based on the dendritic width selected by the user, with the segmentation threshold set to 80% quantile of the neighborhood correlation. Autodetection will identify all the ellipsoid puncta available in the image (e.g., dendritic spines and axonal boutons) based on their spatiotemporal activity, unless the user chooses to specify a dendritic branch on which they wish to detect dendritic spines. In the latter scenario, Autodetection will operate under strict criteria, which allows only those spines that are tangentially proximal (default is within 3 times of width of the dendrite set by the user) to the specified dendritic ROIs to be detected [Fig. 3(b)]. In our three representative image $t$-stacks, autodetection that is unconstrained by the associated dendrite selection had $57.1\%\pm 8.0\%$, $53.5\%\pm 1.0\%$, and $42.9\%\pm 8.0\%$ true positives, false positives, and false negatives detected, respectively, compared to the ground truth set by manual ROI selection by a human expert. Autodetection with the associated dendrite selection had $58.3\%\pm 6.6\%$, $4.6\%\pm 2.9\%$, and $41.7\%\pm 6.6\%$ true positives, false positives, and false negatives detected, respectively (Fig. S1 in the Supplementary Material 1). The higher incidents of false negatives than false positives in the latter approach are expected as the autodetection is set with stringent parameters (threshold set 80% quantile neighborhood correlation within 3 times of dendritic width in pixels). We set these parameters because it makes for a smoother user experience to manually add missing spines than to delete phantom spines from the images with annotations on the GUI. Because the autodetection of spines with a specific dendrite selected restricts its search area within three times of the dendritic width set by the user, users are encouraged to explore and set the dendritic width specific to their dataset that provides the best results.

Fig. 3

Feature detection and ROI segmentation of dendritic compartments and signal extraction. (a) Detection and segmentation process of three example dendritic spines are shown in insets. The initial ROI (1) detected based on seed pixels defined by the user or automatically defined by a particle detection algorithm is segmented by the signal correlation threshold (2) and then constrained by morphological rules through which features beyond the target size range are excluded. An ROI mask is then generated based on the resultant ellipsoid (3). (b) Autodetection will identify all the ellipsoid puncta available in the image (middle) based first on a particle detection algorithm and then segmented by the spatiotemporal activity profiles of the pixels. Users can also specify a long dendritic branch on which to focus the spine detection. For this latter case, only those spines that are proximal to the target dendrite are detected (right). (c) In addition to spine detection, AUTOTUNE can automatically define local shaft subregion ROIs by the subdivision of the user-defined long dendritic shaft into even pieces or via identification of the shaft subregion located in proximity to a spine ROI. (d) Once ROIs are defined, activity transients can be extracted for each ROI. (e) The same ROI map can be applied to multiple source image stacks of the same FOV via cross-session feature alignment. $\theta$ indicates the general angle in which the second stack was rotated to align with the first stack and the master ROIs generated on it.

Unrestricted autodetection can be applied on a batch of multiple registered image stacks but will produce independent ROI map for each. Alternatively, users can opt to use an existing master ROI map, which allows them to realign multiple stacks and apply the single master ROI map to all of them. In contrast, the manual feature detection option will allow users to manually identify long dendritic branches and dendritic spines as well as delete unwanted features including those puncta that were automatically detected using Auto Detection.

In addition to spine detection, AUTOTUNE offers three different methods in which dendritic shafts may be identified [Figs. 3(c) and 3(d)]: (1) user defined identification of a long dendritic shaft (add dendrites), (2) automated subdivision of identified long dendritic branch (subdivide dendrite), and/or (3) identification of dendritic shaft subregion located in physical proximity to an already detected and accepted spine (shaft by spine). The length of each subregion is set by the user in pop-up window. Local dendritic shaft subdivision is especially useful for probing for local dendritic spike-associated changes that are distinct from signals associated with the backpropagating action potentials.^23,40

Once accepted, all the ROIs (long dendritic shaft, local shaft subregions, and spines) are displayed on ROI result box with extracted traces stack plotted in the respective boxes for dendrites and spines [Figs. 3(c) and 3(d)]. After inspecting the feature map, any unwanted features, both dendritic shaft and spines, can be deleted with delete feature function. Once an ROI map is generated, users can use it for subsequently acquired data from the same FOV.

3.4.2.

Cross-session feature alignment

One of the great advantages of fluorescent imaging is the ability to monitor both the dynamics of dendritic and spine morphology and functional activity across multiple imaging sessions. These dynamics include those introduced through natural physiological changes, such as learning and aging as well as others associated with the progression of neurological disorders. Although prior studies have mainly focused on morphological changes, accelerated improvements in GECIs and other indicators have enabled chronic functional monitoring of these subcellular compartments. AUTOTUNE’s feature detection tool allows for batch application of the same master ROI map to image sets acquired over multiple sessions. The user will first use the module, feature detection, on one of the registered image $t$-stacks to generate a master ROI map (e.g., dendritic branches, spines, and local dendritic shafts), which will be then applied to all the registered stacks simultaneously using the batch feature detection module.

Functional images from the same FOV acquired across multiple sessions, days, weeks, or even months apart, can contain larger shifts in ROI locations due to physiological changes over time. For such cross-session comparisons, AUTOTUNE employs the iterative closest point (ICP) algorithm for rigid point cloud-based registration in order to apply a single master ROI map onto each stack.⁵⁴ This process allows users to keep track of the same ROIs in longitudinal studies rather than having to cross-reference multiple maps. Theoretically, the amount of pixel shift in $X$ and $Y$ axes is unlimited [Fig. 3(e)]. However, a realistic limitation applies because features can be lost out of frame with large shifts (Fig. 2). In order to test the versatility of our method, we compared the cross-section alignment performance of AUTOTUNE to the MATLAB built-in function for intensity-based automatic image registration (imregtform⁵⁵). Our batch registration method significantly outperformed its counterpart ($p=0.001$, $t$-test; Fig. 4).

Fig. 4

Cross-session image alignment. (a) Example pre- and postregistration overlays of frame averaged images from two stacks of source images acquired during two separate imaging sessions 1 week apart. (b) Automatically detected key feature maps are first defined for each stack (iter 0), which then are shifted closer together at each iteration (iter 1, 2, …), allowing the program to progressively, but simultaneously correct for displacement and rotation. (c) A simulative quantification showed that ICPs correction used by AUTOTUNE demonstrated superior performance generating significantly less error (MSE, mean square error) compared to a built-in MATLAB function for intensity-based automatic image registration (imregtform). In this test simulation, arbitrary translation and rotation were applied to an image in order to quantify the performance of two registration methods. The fraction of preserved features is a portion of image features that remained within the simulation window after the forced translation and rotation.

3.6.1.

Removal of back-propagating axonal action potential signals

Once the activity traces of the individual ROIs have been extracted, the fourth and final module will allow the users to correlate the trace to any temporally matched parameters, including stimulation features (e.g., orientation of gratings and local optical features in natural images) and other behavioral parameters (e.g., locomotion features, such as speed or location). For mapping dendritic spine calcium signals, the first step that is critically important is to remove any signal contributions from the back-propagating action potentials (bAPs). For this, we have adopted the method by Chen et al.,² in which a linearly scaled version of a long dendritic shaft signals is subtracted from the spine signals^2,19,23 (Fig. 6). Briefly, visually evoked dendritic responses are plotted against those temporally matched responses from each dendritic spine [Fig. 6(b)]. The resultant scatter plot should exhibit two distinct clouds of points; one extending vertically along $y$ axis around $x=0$ and the other extending diagonally. The former indicates spine responses that are independent of the dendritic shaft responses, and the latter represents instances in which both spine and the shaft were activated simultaneously, inferring the presence of bAPs. MATLAB’s robustfit function is then applied to the latter group of points, and the slope of the fit is used as the factor for scaling the raw dendritic shaft signals to be subtracted from the spine signals. This process is performed in a batch where appropriate scale factors are set for all individual identified spine ROIs using the accompanying long dendritic shaft signals. Although the process has been well-adopted by other published works that have advanced our knowledge of dendritic spine physiology,^{2,19,22,23,56} a care must be taken as a simulation has reported potentially biased results created by over- or undersubtraction of bAP contributions.⁵⁷ AUTOTUNE offers immediate visual feedback on the resultant bAP-removed trace, allowing the users to select the largest scale factor (to minimize undersubtraction) for each spine (and local shaft subregion) that does not generate negative deflections (to prevent oversubtraction) [Fig. 6(c)]. Use of the dendritic activity rather than the somatic activity as a bAP activity proxy and exclusion of shaft subregions that exhibit visible nonlinear events (i.e., calcium hotspots from local dendritic spikes) are two important additional considerations that should be taken to minimize the potential biases in the results (Fig. S3 in Supplementary Material 1). bAP removal process can be entirely omitted for signals that do not convey bAPs such as those detected with glutamate sensors.¹ After contaminating bAP signals are appropriately subtracted from the spines, users can perform correlative analysis of the spine signals with a temporally matched stimulus or behavioral parameters of their choosing.

Fig. 6

Removal of bAP for isolating input-specific spine signals. (a) ROI map consisting of a single long dendrite and nine dendritic spines are shown. (b) AUTOTUNE applies robust fit function to the correlated portion of the spine and dendritic activity in order to estimate the scale factor (slope) for the back-propagating action potential (bAP) signal contributing to the spine signals. (c) The dendritic activity trace was scaled down by the factor calculated in (b) (red trace) and subtracted from the two example spines, 1 and 2, to reveal the respective input-specific spine signals.

3.6.2.

Input mapping: orientation tuning analysis of dendritic spine responses

In this study, we offer two example cases in which our program computes correlative relationships between the spine activity and the stimulus or behavioral features, namely, moving direction of oriented black and white gratings (orientation tuning) and locomotive speed, respectively (Figs. 7 and 8).

Fig. 7

Orientation tuning analysis of dendrite and spine responses. (a) AUTOTUNE was used to analyze the orientation tuning of long dendrites and spines. An example dendrite exhibiting sharp orientation preference is shown. The tuning curve was fit with a constrained two-peak Gaussian tuning curve (red). Individual spines also exhibited visually evoked calcium transients and displayed tuned responses of various strengths. (b) Spatially localized dendritic activities were discernible in bAP-removed traces of shaft subregions, potentially indicating incidents of local dendritic spikes.

Fig. 8

Probing for behavioral correlations in spine activity. (a) Schematic showing the trial structure of an example dendritic imaging experiment during a visually guided locomotion task is shown. As a head-fixed mouse walked along the linear virtual treadmill, locomotion speed (instantaneous velocity: blue trace) and dendritic calcium responses (black trace) were recorded. Each trial was considered complete as the mice traveled to the end of the virtual path. The locomotion of the mouse was instantly translated to its position in a virtual track marked with visual cues (closed-loop: red blocks in trial type) or disengaged from the visual feedback (open-loop: blue blocks). AUTOTUNE input mapping/behavioral response module can plot event-triggered averages of the dendritic activity (green shaded area) using a transition in the locomotion speed (e.g., $k$ > value set by the user: red dotted line) and compare between the two trial types. (b) Event triggered averages of calcium transients recorded from five dendrites as mice ($n=2$) accelerated its locomotion ($k>0.5\mathrm{cm}/\mathrm{s}$), comparing between the open- (blue) and closed-loop (red) trials are shown. Time 0 s indicates the initial time of acceleration. (c) Locomotion speed triggered averages of calcium transients from a dendrite, four shaft subregions, and four spines are shown. The $x$ axis is in seconds; $\text{scale bar}=10\mu \mathrm{m}$.

For mapping visual response properties of synaptic inputs arriving at dendritic spines of mouse V1 layer 2/3 pyramidal neurons, AUTOTUNE was used to analyze their orientation tuning. In general, imaged neurons exhibited strong visually evoked activities reflected as global dendritic shaft spiking activities (Fig. 7). An example dendrite showed sharp orientation preference and could be fit with a constrained two-peak Gaussian tuning curve. Individual spines also exhibited visually evoked calcium transients and displayed noticeable tuned responses [Fig. 7(a)]. Local shaft subregion responses were also observed in another example dendrite with evenly subdivided shaft ROIs [Fig. 7(b)]. In this example, spatially localized shaft subregion activities were discernible in bAP-removed traces, potentially indicating incidents of local dendritic spikes.^23,40 These datasets demonstrate the ease and the effectiveness of AUTOTUNE.

3.6.3.

Input mapping: application of dendritic analyses to behavioral settings with multiple stimulus variables

The input mapping module allows us to probe for the circuit-based mechanism of synaptic integration by closely examining potential correlative relationships between dendritic spine responses with complex experimental parameters, such as behavioral parameters with or without stereotyped trial structure (e.g., duration and condition). Here we provide an example test case for how this type of analysis may be performed with AUTOTUNE on dendritic imaging acquired while the mouse traveled in a linear virtual treadmill.

AUTOTUNE enables the comparison of spine activity under different conditions of a select parameter. In this test case demonstration, we analyzed the dendritic activity against the locomotive velocity, focusing on the 4 s window centered around the point at which the animal transitioned from stationary to moving (e.g., instantaneous speed >0.5 cm/s). We used AUTOTUNE to generate event-triggered averaged trace of a long dendritic shaft, shaft subregions, and spines spanning 2 s before and after the transition, which allowed us to explore the potential impact of locomotion on dendritic activities. Moreover, AUTOTUNE allows for trial-structure-based categorization of event-triggered averages for a group comparison [Fig. 8(a)]. The example case compared the locomotive speed-triggered averages of dendritic and spine activities between open-loop and closed-loop trials as the mice navigated a virtual treadmill through a simple visual scene [Figs. 8(b) and 8(c)].

The behavior analysis module is versatile and allows users to define the window size of the trace snippet to inspect, according to their rationale and needs. In addition, for more complex experiments with multiple trial conditions, AUTOTUNE allows users to perform joint analysis on how certain parameter affect neuronal responses under different trial conditions (Fig. 8).