Non-invasive 4D transcranial functional ultrasound and ultrasound localization microscopy for multimodal imaging of neurovascular response

Abstract

A long-standing goal of neuroimaging is the non-invasive volumetric assessment of whole brain function and structure at high spatial and temporal resolutions. Functional ultrasound (fUS) and ultrasound localization microscopy (ULM) are rapidly emerging techniques that promise to bring advanced brain imaging and therapy to the clinic with the safety and low-cost advantages associated with ultrasound. fUS has been used to study cerebral hemodynamics at high temporal resolutions while ULM has been used to study cerebral microvascular structure at high spatial resolutions. These two methods have complementary spatio-temporal characteristics, making them ideally suited for multimodal imaging, but both suffer from limitations associated with transcranial ultrasound imaging. Here, these two methods are combined on the same data acquisition, completely non-invasively, using contrast-enhancements, which solves the dual challenges of sensitivity during transcranial imaging and the ability to implement super-resolution. From this combined approach, the cerebral blood flow, activated brain region, brain connectivity, vessel diameter, and vessel velocity were all calculated from the same data acquisition. During stimulation periods, there was a statistically significant (p<0.0001) increase in cerebral blood flow, diameter, and global velocity, but a decrease in velocity in the activated region. Additionally, the global flow increased (p=0.11) and connectivity decreased (24.7%) when compared to baseline. This multimodal approach allows for the study of the relationship between cerebral hemodynamics (30 ms resolution) and the microvasculature (14.6 m resolution) using one ultrasound scan.

Introduction

Understanding the relationship between measurements of brain function and behavior remains one of the most significant challenges in neuroscience. The brain is an electrical network of neurons and synapses that generates complex activity patterns to control our perceptions, thoughts, and movements. When active, these neural circuits interact to perform numerous brain functions from interpreting sensory input to performing tasks. Understanding brain circuits and their relationship to function is a fundamental objective in neuroscience. However, describing these relationships requires measuring the whole-brain neurofunctional response to stimuli with high spatial and temporal resolution. Different non-invasive whole brain neuroimaging modalities have been developed for this application, including functional magnetic resonance imaging (fMRI), which has a spatial resolution of 1–2 mm and temporal resolution of 1 s, positron emission tomography (PET), which has a spatial resolution of 4 mm and a temporal resolution of 5 s–5 min, and single-photon emission computed tomography (SPECT), with a spatial resolution of 4 mm and temporal resolution of 5–15 min¹. There are neuroimaging techniques with high temporal resolution, but they compromise on spatial resolution. Electroencephalography (EEG) has a high temporal resolution of <1 ms but cannot be used to resolve the exact location of fine responses given its poor spatial resolution of 0.7–1 cm². Magnetoencephalography (MEG) has the same <1 ms temporal resolution with improved spatial resolutions of 2–3mm but it cannot resolve data coming from non-tangential directions and can only reach shallow brain regions^2,3. The spatio-temporal scales of brain function, however, are sub-mm and ms. Optical imaging methods have achieved high spatio-temporal resolutions but are limited by depth of field (500–600 m) and required invasive procedures (skull or scalp removal) to image the brain^4,5. PET has been combined with computed tomography (CT)⁶and magnetic resonance (MR) imaging⁷ to combine functional and structural information, but these techniques are still limited in temporal resolution. To further our understanding of the brain including when and how it responds to events, a non-invasive neuroimaging technique with large imaging depths and both a high spatial and temporal resolution is needed.

Functional ultrasound imaging (fUS) can detect changes in blood flow with spatial resolutions of 200 m and temporal resolutions of 20 ms⁸. This provides a technique to image the brain with a high spatio-temporal resolution. Two-dimensional fUS has been performed on neonates⁹, and on non-human primates and adults with a craniectomy^10–12. Transcranial 2D fUS has been performed in rats¹³and mice¹⁴during stimulatory events. This technique records changes in blood flow for a brain slice. However, the functional response is often complex and involves many different brain regions within the brain circuit. To create a complete picture of the brain response through time, 3D images are required. There are ways to generate volumes using 2D imaging using mechanical scanning, which has been used to detect hemodynamic changes during stroke in awake rodents, but this requires long scan times and cannot resolve out-of-plane vessels^15–17. This means that to accurately characterize the brain, volumetric imaging is required. While not yet whole brain, fUS has been used to perform volumetric functional imaging^8,18,19. Very recent applications of volumetric ultrasound to perform functional imaging with a matrix array in rodents required a craniectomy^8,18. A row-column array has been used to perform fUS transcranially, with scalp removal, but sacrificed spatiotemporal resolution to build brain volumes¹⁹. Completely non-invasive, transcranial volumetric fUS, with no scalp removal, has not yet been demonstrated to image the brain response during stimulation with a continuous temporal resolution. Additionally, current volumetric functional imaging techniques – like fMRI, which measures the average diffusion within a pixel, and fUS, which measures the power signal within pixels – do not provide information about the surrounding vasculature. To identify individual components of the microvasculature that are responsible for neurovascular coupling, anatomical information about the location of the functional signal is needed.

Imaging brain vasculature is challenging due to the small size of vessels, which have diameters on the scale of microns, the speed of blood flow, varying between 1 mm/s to 100 mm/s, and challenges introduced by the skull²⁰. Both ex vivo– including perfusion micro computed tomography and clearing optical microscopy^21–24 – and invasive in vivo– including two-photon microscopy and near infrared spectroscopy^25–30 – techniques have been used to image changes in vasculature. These techniques suffer from a combination of drawbacks including limited penetration, small field of view, and invasiveness. Changes in vasculature may not be local and compensated by collateral circulation, so these methods are limited in their ability to quantify in vivowhole brain changes. To overcome these challenges to imaging the microvasculature, ultrasound can be used. Ultrasound localization microscopy (ULM) is an ultrasound technique that was first developed in optics to resolve structures beyond the physical resolution limit of the imaging system³¹. To create super-resolution images of the microvasculature, this technique has been used to localize microbubble contrast agents (MBs) circulating in different organs in rodents^32–38, including in the brain through either a thinned skull³⁹or with a craniectomy⁴⁰. This method has also been performed in humans^41,42, including in the brain⁴³. Volumetric ULM has recently been performed in rats and mice with a spatial resolution of 20–31 m and a flow rate between 2 mm/s to 100 mm/s^15,44,45, making it a critical technique for the study of brain microvasculature, with some studies examining the whole brain through stitching together multiple volumes⁴⁶. Recently, it has been shown that non-invasive volumetric ULM images in rats and mice can be created without the need for a craniectomy⁴⁷. ULM can be used to provide the anatomical context that functional imaging techniques are missing.

Functional ULM has been demonstrated in anesthetized rats with craniectomy to measure the changes in blood flow due to stimuli in the microvasculature⁴⁸but this technique decreases temporal resolution. These techniques provide some information about both brain structure and function, but they have only been performed in 2D with a craniectomy. In addition, this technique requires the combination of data during ten simulation events to prevent sparse representation of the vasculature, limiting its effectiveness to study individual or non-uniform stimulation events and increasing the acquisition time required to create these images. Several other ULM techniques have been used to study microbubble flow including Dynamic ULM to measure cerebral pulsatility and sensing ULM to detect microbubbles with different flow behaviors, like glomeruli in the kidney^49–51. However, to fully investigate brain circuits, both structural and functional information is required. A recent study has created both power Doppler and super resolution images in the kidney from one dataset⁵². However, this technique was in 2D and was not applied to the brain.

Here, both volumetric ULM and fUS were performed non-invasively in the same acquisition through the intact skull and scalp of a rat using microbubbles, for the investigation of stimulatory events. This implementation is innovative because it is (a) completely non-invasive, without skull thinning or scalp removal, (b) combines two techniques that have not previously been performed on the same acquisition in the brain volume, (c) allows investigation of individual stimulatory events and (d) generates volumetric movies of the individual functional responses to a stimulus through time as well as a super-resolved volumetric image of the underlying cerebral vascular map, providing a detailed microvascular context for the location of the functional response, without event averaging required. These new capabilities can enable the use of volumetric ultrasound imaging as a low cost, fast, portable modality for recurrent functional brain studies, including the investigation of the progression of different neurological and psychiatric diseases, behavior studies during learning, and the effects of neuromodulation or ablation.

Results

Volumetric Ultrasound Localization Microscopy

To create volumetric ultrasound localization microscopy (ULM) images of the brain microvasculature, an 8–9 week old Fischer344 rat (Charles River, Wilmington, MA) was imaged using the techniques described in McCall et al.⁴⁷. A 1024-channel Verasonics (Verasonics, WA) volumetric Vantage system was used with a Vermon (Vermon S.A., Tours, France) 32x32 matrix array transducer to transcranially image the brain of an anesthetized and depilated rat through an intact skull and scalp. Microbubbles were infused through the tail vein to act as a contrast agent. A diagram of the setup is shown in Fig. 1. The acquisition lasted for 320 s where a vibrotactile stimulus was applied to the hind paw in a 30 s on, 40 s off pattern, repeated five times, following a 90 s baseline acquisition. The data was beamformed and Singular Value Decomposition (SVD) filtering was applied, followed by microbubble localization and tracking. Volumetric images of the microvasculature were created using 3dSlicer⁵³ The resolution of the ULM image created using the 320 s dataset with stimulation is 14.63 m, as calculated using Fourier Shell Correlation (FSC). The corresponding FSC plot is shown in Supplementary Figure 1 (b).

Fig. 1.

(a) Diagram of the experimental setup. A stereotactically-fixed rat being infused with contrast agent was imaged with a Vermon 8MHz matrix array being controlled by a 1024-Channel Verasonics 3D imaging platform. During imaging, the Verasonics triggers an arbitrary function generator that mechanically stimulated the hind-paw, activating the S1HL area, or the contralateral somatosensory region, of the rat brain. Created with BioRender.com. (b) Experiment acquisition timeline demonstrating which portions of the acquisition were used for which image and showing when stimulation was triggered.

Images of the cerebral microvasculature were created for the 8–9 week-old rat across the acquisition to create an anatomical map for the functional response. The coronal and sagittal views of the resulting volumetric ULM image are shown in Fig. 2 (e) and (f). The intermediate processing steps, including b-mode slices in time (Fig. 2 (a) and (b)) and the maximum intensity projections (MIP) through elevation (Fig. 2 (c) and (d)) of the Singular Value Decomposition (SVD) filtered b-mode are shown. With this method, a spatial resolution of 14.27 m was achieved for the 320 s dataset with stimulation, as measured by Fourier Shell Correlation, creating our anatomical map of the microvasculature. The velocity at the center of each vessel was calculated with resulting blood velocities of −76–95 mm/s. The velocities across the volume are shown in Fig. 2(g) and (h). The resulting ULM image for the 8–9 week-old rat was aligned with an atlas to segment the brain into its different brain regions⁵⁴.

Fig. 2.

A beamformed coronal (a) and sagittal (b) slice. Maximum Intensity Projection of the SVD filtered data of 5 seconds during baseline with the coronal (c) and sagittal (d) dimension collapsed. Volumetric ULM image with the coronal (e) and sagittal (f) views. Volumetric plot of microbubble velocities with the coronal (g) and sagittal (h) views. Blue, or negative, velocities represent those moving away from the transducer and red, or positive, velocities represent those moving towards the transducer.

Data acquisition and beamforming for high speed ULM

ULM imaging is a highly capable tool for visualizing the vasculature, but it still faces challenges in clinical translation due to the requirement of long acquisition times. Shorter acquisition times would aid its translation to the clinic since they reduce motion artifacts and imaging time. The ULM results were compared at four different acquisition lengths of 10 s (Fig. 3 (a)), 30 s (Fig. 3 (b)), 60 s (Fig. 3 (c)), and 90 s (Fig. 3 (d)), on the baseline dataset to avoid any effects from the stimulation on/off periods. Across the four time points, the resolution, measured using FSC, improved from 14.27 m to 14.00 m as the time increased. In addition, the sparsity of the vessels decreased as network filling improved. For the 10 s acquisition case, it took 20 minutes to generate a final super resolution image on our GPU workstations, including a custom delay-and-sum beamformer, programmed in CUDA (Nvidia Corporation, Santa Clara, CA), for parallelized graphics processing unit (GPU) processing (1 min), SVD filtering (18 min), and bubble localization and tracking (1 min), all performed on the GPUs. For the 30 s case it took 58 minutes, 118 minutes for the 60 s case, and 180 minutes for the 90 s case. For a 1,300 mm³ volume, in this case encompassing 1,373,328 voxels, this custom beamformer completed beamforming a full 90 s dataset (45,000 frames) in 11.35 minutes. This beamformer drastically improves processing speeds for volumetric data. While it is not yet real-time, future computational improvements and better GPU implementations are moving volumetric data processing closer to real-time.

Fig. 3.

Volumetric ULM images created using 10 s (a), 30 s (b), 60 s (c), and 90 s (d) of data, plotted on the same colorbar axis.

Volumetric functional ultrasound imaging

In addition to the ULM vascular maps, functional ultrasound imaging (fUS) was performed. To create volumetric functional movies through time, power Doppler processing⁸ was performed on the same dataset as a measure of cerebral blood flow (CBF). Analysis was focused on the contralateral somatosensory cortex to the paw with stimulation applied, in this case the left somatosensory cortex, which is the region of the brain that interprets vibrotactile information. Elevational MIPs of the power Doppler intensity over 0.2 s are shown in Fig. 4. The “on” and “off” MIPs are both individually normalized. The box is highlighting the contralateral somatosensory cortex region, with a gray box representing when stimulation is off (Fig. 4 (a, c) and a green box representing when stimulation is on (Fig. 4 (b, d). These images have a temporal resolution of 30 ms due to our sliding window step size.

Fig. 4.

Power Doppler MIPs showing the power Doppler intensity (A.U.) over 0.2 s in time for when stimulation is off (a, c) and on (b, d). The coronal view, taken from the maximum through elevation, is shown in (a, b) and the sagittal view, taken from the maximum in the lateral dimension, is shown in (c, d). The box highlights the contralateral somatosensory cortex region.

Analysis of stimulation response

To analyze the CBF during stimulation, the individual stimulation plots for the right cerebral cortex, left cerebral cortex, and two control regions are shown in the left column of Fig. 5 (a), (b), (c), and (d), respectively. The green portions represent when stimulation is on and the red portions represent when it is off with a 30 s on, 40 s off stimulation pattern. The right column shows a graph of the total mean power Doppler signal in each on and off period, including a mean bar and a confidence interval of 95, with five on and four off periods observed over the course of 320 s. Based on a two-way ANOVA with a Bonferroni correction comparing CBF in brain regions during on and off periods across the five cycles, all brain regions had significant differences (p<0.01) in CBF between on and off periods except for the left control (Supplementary Table 1). However, the difference between on and off for the right control was a negative difference in the mean value, indicating that the CBF in the right control was significantly higher when stimulation was off than when it was on. Between brain regions, all groups had a significant difference in CBF except between the left and right control regions when stimulation was on and when stimulation was off. There were no significant differences in CBF between cycles (Supplementary Table 2).

Fig. 5.

Individual stimulation plots based on the mean power Doppler intensity for the right cerebral cortex (a), left cerebral cortex (b), right control region (c), and left control region (d). The plots are baseline corrected based on the five seconds preceding each stimulation event and a moving average is applied for visualization. For the individual plots, the red locations represent when stimulation is off, and green represents when it is on. Next to the stimulation plots are plots of the mean power Doppler value in each of the on and off periods with a 95 confidence interval and a mean bar.

When the cross-correlation between the stimulation profile and the power Doppler intensity stimulation plots was performed, the delay in response, based on when the signal has the maximum correlation with the stimulation train, was found to be 1.56 s for the right somatosensory cortex from the onset of stimulation. The left cortex, right control region, and left control region responses were behind the right cortex by 4.59 s, 12.57 s, and 13.32 s, respectively.

Another measure of changes in brain activity during stimulation is the connectivity matrix, which here measures the covariance between brain regions through time. Typically, connectivity is high at baseline since the whole brain is spontaneously firing at the same time. Then, when stimulation is applied, these brain regions decorrelate, leaving high correlations in the stimulated circuit⁵⁵. For hind paw stimulation, the brain circuit includes the S1, S2, and the thalamus, which should have high connectivity. Regions outside this brain region should have low connectivity⁵⁶. For this study, the connectivity matrices across the entire brain volume, before and during stimulation, are shown in Fig. 6. The baseline case has an overall higher connectivity, with a mean connectivity value across the brain of 0.80 ± 0.11 compared to 0.60 ± 0.19 for the case with stimulation. For the baseline case, there is high connectivity throughout. During stimulation, when comparing the cortex, where activation due to stimulation occurs, with the other brain regions, the highest connectivity is seem with brain regions involved in interpreting vibrotactile stimulus, like the hindbrain (0.84), midbrain (0.71), thalamus (0.69), and cerebellum (0.86). The region with the lowest connectivity to the cortex is the globus palladis (0.42). However, there is high connectivity with the caudoputamen (0.79), which is next to the globus palladis in the basal ganglia.

Fig. 6.

The connectivity matrices calculated using the squared Pearson correlation during baselines (a) and vibrotactile stimulation (b) across 16 selected regions of interest are shown.

Combined functional and super resolution images

The ULM images shown have a high spatial resolution of 14 m and the fUS images have a high spatial temporal resolution of 30 ms. Since these techniques were performed on the same acquisition, they are registered with one another and can be overlaid, providing information about which vessels are experiencing a functional response to the vibrotactile stimulus. The combined fUS and ULM image are shown in Fig. 7, with the fUS shown in blue, and the ULM, shown in red. This allows us to determine the exact vessel that is experiencing activation during our vibrotactile stimulation. In Fig. 7, the gray box represents a time point when stimulation is off, and green when stimulation is on, with the box highlighting the S1HL region. Between on and off periods, some of the larger vessels have a high power Doppler signal in both cases. However, when stimulation is on, there is increased power Doppler signal in the highlighted region. A movie of the CBF over 3 s of time is shown in Supplementary Movie 1. The combined fUS and ULM images and movie for a younger, 3–4 week-old rat, are shown in Supplementary Figure 2 and Supplementary Movie 2.

Fig. 7.

Combined fUS and ULM images where a fUS image at one time point (blue colormap) is plotted over the ULM image (red colormap) for when stimulation is off (left column) and on (right column). The volume from both the coronal (first row) and sagittal (second row) views are shown.

In previous studies, a concentration of 3MB/mL/g of the animal has been used for volumetric ULM images⁴⁷. However, contrast enhanced fUS imaging requires high microbubble concentrations to achieve a more constant bubble flow. Due to this, two datasets were acquired on the same animal and in the same position, one with a concentration of 3 MB/mL/g and one with 3 MB/mL. Then, the resulting combined ULM and fUS images were compared in three cases. One case where both were performed on the higher concentration, one where both were performed on the lower concentration, and one where the ULM was performed on the lower concentration and the fUS was performed on the higher concentration. The resulting images from one fUS time point are shown in Fig. 8. Between the images, the ULM and fUS images were plotted on the same two scales. For the ULM tracked data across 320 s, the resolution decreased from 12.83 m to 14.63 m when increasing the concentration; however, the number of total microbubbles localized increased from 2,955,573 to 3,069,276. The FSC resolution plots for the lower and higher concentrations are shown in Supplementary Figure 1. The total power Doppler signal also decreased 180 at the lower concentration. When calculating the cross correlation with the stimulation train, the maximum correlation in the right somatosensory cortex fell from 0.8 to 0.6 and from 0.35 to 0.28 in the left between the higher and lower concentrations. The control regions had similar maximum correlations of 0.32 and 0.13 in the right and left control regions for the higher concentration, and 0.33 and 0.13 for the lower concentration.

Fig. 8.

Combined fUS and ULM images for one time point using different injected concentrations. (a) The combined fUS and ULM image where they are both performed on the acquisition with a microbubble concentration of 3 MB/mL/g. (b) The combined fUS and ULM image where both are performed on the acquisition with a microbubble concentration of 3 MB/mL. (c) The combined fUS and ULM image where the ULM is calculated from the dataset with the lower concentration of 3 MB/mL/g and the fUS is calculated from the dataset with the higher concentration of 3 MB/mL. All ULM images and all fUS images are plotted using the same colorbar axes.

Vascular response to stimulation

From the combined image, a fUS activation map can be calculated using the t-statistic of every voxel using a generalized linear model (glm) between the power Doppler signal in time and the stimulation train. This activation map can be used to find regions of activation on the underlying structural images. The activation map overlaid on the vasculature is shown in Fig. 9, where the minimum cut-off t-statistic value was chosen to be one fourth of the maximum value. When comparing the values per vessel across the brain volume for diameter, velocity, and blood flow, there is a significantly higher velocity during stimulation across the brain (p<0.0001). Neither diameter nor blood flow are significantly different globally. When comparing the activated region during on and off periods, the diameter across the activated region was significantly higher during vibrotactile stimulation (p<0.0001), whereas the velocity was significantly lower during stimulation (p<0.0001). There was no significant difference in blood flow (p=0.11) but there was a higher mean flow value of 1.78 L/min while stimulation was on compared to 1.43 L/min whilst off.

Fig. 9.

The activation map, shown in purple, overlaid on the ULM image, shown in red, revealing activation in the S1HL. The activation was calculated using GLMs between the whole power Doppler dataset and the stimulation train. The activation is reported by the t-statistic value with the minimum cutoff value set to 1/4 of the maximum value.

Discussion

Here, completely non-invasive volumetric fUS and ULM have been performed on the same dataset, combining the high spatial and temporal resolutions of each technique, providing a way to study both structural and functional information, like blood vessel velocity, and brain region connectivity, using one acquisition. Hind-paw vibrotactile stimulation is an established method to measure functional brain response to activation, with previous methods showing that upon stimulation, there is an increase in functional activity in the contralateral somatosensory cortex^57–59. This stimulus was used to benchmark our imaging approach of using completely non-invasive fUS of Fischer344 rats.

In this study differences could be detected in functional brain response between periods with and without stimulation applied, seeing increased activity across the brain, with highly significant differences in the somatosensory cortices. There was a significant increase in CBF in the somatosensory cortex during hind-paw vibrotactile stimulation based on the power Doppler intensities in those regions. The right and left somatosensory cortex had a highly significant difference in CBF between on and off periods (p<0.0001), with on periods having a higher CBF than off periods. The right somatosensory cortex is responsible for interpreting signal from the left hind paw. In the brain, it is common to see a weaker, delayed contralateral response in the corresponding brain region on the opposite side of the brain^60,61, corresponding to the left somatosensory cortex in this case. This is also seen in the activation map in Fig. 4(c, f). The right control region also had a significant difference in CBF between on and off periods (p<0.0001); however, this region saw a significantly higher CBF during off periods, instead of during on periods. This could be due to a delayed response in that region or due to the large redirection of blood to the contralateral somatosensory region. The left control region did not see a significant difference in CBF between on and off periods (p=0.03). Between cycles, there were no significant differences in CBF. There is a slight decrease in magnitude of response in the right somatosensory cortex between the five trials in the smoothed version that is likely caused by a decrease in microbubble concentration through time, but based on our two way ANOVA, this difference is not significant in the non-smoothed data. This could be mitigated in the future through the use of a higher infusion rate to compensate for the rate at which the animals metabolizes the microbubbles. There are some regions outside of the S1HL with high CBF values during stimulation; however, it has been shown that the brain circuit involved in interpreting sensory input invovles S1, S2, and the thalamus⁵⁶. Other regions that had a high CBF in the MIPs did not necessarily follow the stimulation pattern, as they were eliminated in the activation maps, leaving the S1HL and deeper regions including the thalamus.

The activated area inside the S1 region was smaller than in previous fUS studies. However, previous studies have used whisker stimulation, whereas here, hind paw stimulation was used. It has been shown that using electrical stimulation on a paw digit, one activated vessel can be detected⁶². Furthermore, when using electrical stimulation at 24 Hz, there is approximately a 10x smaller activation area than when performing whisker stimulation using air puffs at 10 Hz⁶³. There are also some other, deeper regions, that show activation. However, the majority of this activation is in the thalamus, which is part of the somatosensory brain circuit⁵⁶. Recent studies using 15.2 T fMRI have shown activation in the thalamus and S2 regions that could not be detected with 9.4 T fMRI during forepaw electrical stimulation⁶⁴.

To determine the delay in response of each region behind the onset of stimulation, the correlation was calculated. the right somatosensory cortex had the highest CBF per voxel and the smallest delay after stimulation onset of 1.56 s. The left somatosensory cortex had a 35 lower amplitude and a larger delay between stimulation onset and point highest correlation of 6.15 s, corresponding with a delayed, opposite brain response. The two control regions had very large delays in their response time. This aligns with the significantly higher CBF during off periods in the right control region. The left region did not have a significant difference, but did also have higher mean value during off than on periods. The appearance of a pattern could be due to natural fluctuation, changed in CBF due to the stimulation, or the introduction of noise from our baseline subtraction and averaging.

There was a large difference in the connectivity between the case with and without stimulation. For the case without stimulation, a high average connectivity of 0.80 was seen throughout most of the brain due to the lack of any external stimuli. Then, when vibrotactile stimulation was applied, the average connectivity decreased by 24.7to 0.60. During baseline, the brain experiences spontaneous activation, correlated across the brain volume. However, when stimulation is applied, those brain regions decorrelate⁵⁵. Since different stimulation activates different brain circuits, areas within the same brain circuit will be connected to each other, responding to the on and off periods of the stimulation. In this case, that involves regions in the cortex and thalamus⁵⁶. Additionally, connectivity also remains high in areas responsible for regulation and the relaying of information like the midbrain and hindbrain. Previous studies have shown high connectivity in the thalamus, caudoputamen, cortex, hindbrain, and midbrain regions due to hind paw stimulation, which is also seen here^8,65. The globus pallidus experiences the lowest connectivity during stimulation across the brain volume (0.46 ± 0.06). Typically the connectivity is high between the globus pallidus and the cortical regions^66,67. However, this region is next to the caudoputamen, another region within the basal ganglia, which does have high connectivity across the volume (0.78 ± 0.08). The differences could be related anesthesia, where previous studies have shown that the basal ganglia, including the globus pallidus, is sensitive to brain state under anesthesia⁶⁶. However, this is also a small region. The connectivity per region was calculated based on our brain segmentation determined by our atlas registration. Since this registration was performed manually on an atlas without vasculature, there is some intrinsic error that comes from this brain segmentation.

The older, 8–9 week-old, rat used in this study had reached the brain volume and skull thickness of an adult rat; however, it had not reached full neurological maturity, which occurs around 12 weeks of age⁶⁸. The younger, 3–4 week-old, rat, shown in Supplementary Figure 2, had not yet reached adult skull thickness, which is reached one month after birth⁶⁸. The younger rat has less CBF outside of the somatosensory region, with a mean power Doppler intensity that is 27% higher than the mean value in the older rat, as well as an improved SR resolution by 1.8 m. This is likely partially due to a difference in skull softness due to the age of the rats, making it harder to transmit ultrasound through the older rat skull. There could be some confounding factors due to their neurological maturities. However, previous studies have found that BOLD levels in the somatosensory cortex due to stimulation reach levels similar to those in adults by 30 days after birth^69,70. While the main difference between the datasets is likely due to the skull thickness, there could also be some effects due to the fact that the MB concentration was not changed for the younger rat, which was smaller and had less blood volume. In addition, the anesthesia was manually tuned based on observed respiration rate, which could introduce variability. Even though the images in the older rat show our capability to perform this technique non-invasively in adult rodents, the images of the younger rat show how clear this technique can be for functional imaging, augmenting its exciting potential as a non-invasive, high spatio-temporal resolution imaging technique for the neurofunctional imaging field. The younger and older rat were also both manually registered to the same brain atlas without accounting for differences in age; however, the atlas was scaled for their different brain sizes accordingly. This paper shows better resolution, as measured by FSC, and higher velocities than previous studies using non-invasive 3D ULM⁴⁷. This is likely due to the implementation of an improved transducer calibration⁷¹. Previous 2D studies have detected an 11 m resolution of ULM images in the brain through a cranial window⁷².

This fUS technique uses microbubbles, unlike other invasive fUS techniques including existing volumetric fUS imaging techniques⁸. However, this use of microbubbles allows us to detect functional signal without the need for skull or scalp removal, allowing us to perform non-invasive transcranial imaging. In addition, the use of microbubbles allows us to perform both ULM and fUS on the same dataset, combining the high spatial resolution of ULM with the high temporal resolution of fUS. In fUS and other functional modalities, it can be hard to determine the exact location of brain activity without overlaying the results on an image using another imaging modality that can perform anatomical neuroimaging. Here, these were combined on the same dataset, from a single ultrasound imaging system, allowing for the study of functional activity with its underlying anatomical context.

To provide the anatomical context, the velocities and diameters were calculated using the extracted centerlines. However, vessel segmentation, when dependent on image intensity, can result in disjointed segments. Here the same segmentation was used, performed over the whole dataset, was used to compare values during on and off periods, using any segments within the activated region. Even though this technique uses MBs, this approach does not require averaging to detect the functional response, unlike other ULM based methods. Other functional techniques that use microbubbles include fULM⁴⁸. This technique utilizes ULM with the averaging of time events to increase spatial resolution. However, this is at the detriment to temporal resolution, requiring multiple stimulation events to achieve adequate network filling for a movie of the functional brain response. This technique provides information about the vascular structure and average flow but is not used for connectivity of brain circuits. Our combined fUS and ULM technique allows us to maintain both high temporal and spatial resolution, while visualizing both the microvascular structure and the functional response through time. This technique can be used to detect differences between on and off periods of one stimulation cycle, at a temporal resolution of 30 ms, compared to averaging ten cycles of that same length. However, for whole brain functional circuit analysis to be performed with the high resolutions of ultrasound imaging, whole brain imaging techniques will need to be developed. This could involve larger transducers or the stitching together of different volumes.

ULM images were also created with low acquisition times under 2 minutes, but there is a trade off between time and vascular network filling and resolution⁷³. Even though the resolution and network filling improves with longer acquisition times, our image for 10 s of data shows promise for high quality images of the microvasculature with small acquisition times. Additionally, it currently requires 20 minutes to create a super resolution image using 10 s of data and 180 minutes for 90 s of data using our current hardware and implementation. This combination of imaging and processing time is not only much faster than other vascular imaging techniques, but shows the feasibility of real-time super resolution and functional volumetric ultrasound imaging. When comparing the ULM images created using 10 s, 30 s, 60 s, and 90 s of baseline data, the resolution improved from 14.27 m to 14.0 m. However, the 60 s ULM image had a resolution lower than that of the 90 s image of 13.95 m. This could be due to similar network filling between the two datasets. However, these both have a better resolution than the ULM image created on the 320 s simulation dataset, with a resolution of 14.63 m. It has been shown the FSC resolution measurements can be affected by undersampled data or data with large artifacts⁷². It is possible that some artifacts are introduced due to erroneous tracks in the longer datasets that are not present in the shorter ones.

For the combined images, there is an additional trade off between ULM resolution and fUS quality. For accurate fUS images, a large number of microbubbles is needed to accurately maintain the transient functional responses in the brain. Therefore, a decrease in power Doppler signal of 35 reduces the fUS quality, shown in the lower correlation between the power Doppler signal and the stimulation trains, ranging between a 20 and 60 decrease in correlation for our four regions. When comparing the images at different concentrations, the resolution only decreased by 1.8 m for the ULM images when using the higher concentration. This supports the conclusion that the higher concentration can also be used for super resolution images. The ULM images also appear to have a higher vessel density at the higher concentrations, which could be due to more microbubbles reaching more vessels, particularly smaller ones. The total number of microbubbles did not increase significantly at the higher concentration, but this could be due to the decreased localization of microbubbles in large vessels due to a higher rate of microbubble overlap and the increased localization of microbubbles in small vessels. Since there is only a slight reduction in ULM resolution at the higher concentration level, the ability to perform both ULM and fUS on the same dataset outweighs the potential benefits of performing ULM on a separate, lower concentration, requiring another acquisition. This increases acquisition time, processing time, and introduces the potential for differences in microbubble syringe concentrations and slight shifts in position between the two datasets, removing their inherent registration. This also removes the ability to perform connectivity and structural analysis on one dataset with unique stimulation events. In future studies, the ULM resolution could be improved on the higher concentration datasets using different approaches to localize more bubbles at higher concentrations, like deep learning approaches⁷⁴.

It was demonstrated that 4D fUS and ULM can be performed simultaneously and non-invasively, allowing for temporal analysis using the fUS dataset with a resolution of 30 ms, and spatial analysis using the corresponding ULM dataset with resolution of 14.6 m, drawing from the complementary strengths of each separate modality. Non-invasive 4D fUS imaging was performed on anesthetized rats during hind paw stimulation, allowing for changes in blood flow due to stimulation to be detected and demonstrates the feasibility of using ultrasound as a non-invasive modality to study transient hemodynamic responses in the brain. Multimodal fUS and ULM vascular maps can be generated, providing quantitative assessment of the anatomy and vascular dynamics, which cannot be done in a single acquisition using other imaging modalities, bringing us closer to the goal of whole brain functional microscopy. Using this technique, functional metrics like connectivity can be combined with stuctural metrics like vessel diameter and velocity to provide a comprehensive understanding of the brain. Due to the non-invasive nature of this technique, imaging can be performed recurrently to study changes in structure and function over time. This can be applied to study both changes in brain function and structure through time for many different neurological conditions including stroke and Alzheimer’s Disease as the disease progresses. Providing a better understanding for how the brain changes due to these disorders can lead to improved research into early diagnosis, treatment, and recovery.

Methods

Experimental setup

Data was acquired on an 8–9 week-old Fischer344 rat (Charles River, Wilmington, MA). Anesthesia was induced with 5 isoflurane gas carried by medical air with anesthesia levels maintained at 2isoflurane throughout the experiment. The head was depilated and placed in a stereotaxic frame (Stoelting Co., Wood Dale, Illinois, USA) and the tail vein was catheterized. A Vermon 32x32 matrix array ultrasound transducer (Vermon S.A., Tours, France) was positioned over the depilated head on top of echographic gel for coupling. A custom vibrotactile stimulation device was positioned on the left hind paw of the rat. The stimulus was controlled with a Tektronix AFG3101C single channel arbitrary function generator (Tektronix, Beaverton, OR, USA). The stimulation and image acquisition were both triggered with a Verasonics (Verasonics, WA) Vantage system to synchronize the timing of the stimulus and imaging. The stimulus device vibrated at a frequency of 25Hz in a 30 s on, 40 s off pattern, repeating five times, following an initial 5 s off. This vibrotactile stimulation of the left hind paw is interpreted by the contralateral somatosensory cortex of the rat brain^58,75,76. Due to this, most of the analysis was performed on the Somatosensory cortex. The experiment involved one 90 s baseline acquisition followed by a 320 s acquisition with stimulation. In-house fabricated microbubbles with an average diameter of 1 micron were infused into the animal through the catheterized tail vein⁷⁷. A microbubble concentration of 3 MB/mL was infused at 15 L/min throughout the experiment. All experimental procedures on rats were performed in accordance with the National Institutes of Health guidelines for animal research, ARRIVE guidelines, and with approval from the University of North Carolina at Chapel Hill Institutional Animal Care and Use Committee (IACUC).

Imaging sequence and beamforming

Data acquisition was performed using a 32×32 Vermon matrix probe transmitting at 7.81 MHz with a custom, asynchronous streaming sequence on a 1024-channel Verasonics (Verasonics, WA) volumetric system⁴⁷. Five angled plane waves, starting with a 0 transmit and then rotated ± 3 elevationally or laterally, were used at a continuous volumetric frame rate of 500 volumes/s, where a 1cmx1cmx1.3cm volume was acquired. A one cycle pulse was emitted with a duty cycle of 67% and a mechanical index of 0.5. Each system of the volumetric scanner was outfitted with an 8 TB raid-0 drive, with a transfer to host time of 750 MB/s and saving rate of 870 MB/s. Volumetric beamforming was performed using a custom delay and sum beamformer written in CUDA (Nvidia Corporation, Santa Clara, CA), allowing for high speed parallelized volumetric beamforming. Data was beamformed on a grid of /2 /2 /2, where =1.9744m. A local graphics processing units (GPU) workstation with four Nvidia (Nvidia Corporation, Santa Clara, CA) RTX-3090 and 48 cpu threads running linux kernel 5.15.0–86-generic was used. All data was processed in MATLAB (The Mathworks, Inc., Natick, MA) and visualized using either MATLAB, for 2D images, or 3dSlicer⁵³, for 3D images.

Volumetric Ultrasound Localization Microscopy

To create highly detailed images of the microvasculature, volumetric ULM was performed. This technique localizes and tracks microbubbles in the blood stream to create images of the microvasculature. To do this, the beamformed data was filtered using 4D Singular Value Decomposition (SVD)⁷⁸, where 30% of singular values were removed, calculated every 200 frames. This filters out portions of the data that spatio-temporally decorrelate slowly, like the skull and stationary tissue, and preserves the signals that spatio-temporally decorrelate quickly, such as microbubbles flowing through the vasculature. The microbubble signal was localized using a weighted centroid approach previously described in McCall et al.⁴⁷. and tracked using the Hungarian Algorithm (simpletracker)⁷⁹to resolve the location of blood vessels due to microbubble flow. This tracking provides information about the location of each bubble through time, allowing for the extraction of velocity and direction. Only tracks that crossed at least three frames and did not travel more than fifteen voxels or five frames between consecutive localizations were included. The image resolution was quantified using Fourier shell correlation with a 1/2-bit threshold, which has been adapted for use in one image in Koho et al.⁸⁰. Volumetric renderings were created using the maximum intensity projection rendering technique with no gradient or scalar opacity thresholds in 3dslicer⁵³with no masking. The 90 s baseline dataset was used to create the images comparing resolution across different acquisition times. The 320 s dataset with stimulation was used in the overlaid ULM and functional images. The ULM images were skeletonized by first applying a threshold, determined using Otsu’s method, to binarize the image⁸¹. The vessel centerlines were extracted from the binarized image using a GPU-based centerline extraction technique⁸². To calculate the diameter, the shortest distance between each voxel along the centerline and the edge of the binarized image was found. Then, the average velocity at each point along the centerline was calculated based on the average velocity of the microbubble tracks going through that location. Assuming parabolic laminar flow, the velocity at the centerline should be the maximum value, where the average velocity is half of this maximum value. The flow was calculated by multiplying the surface area, based on the diameter at each centerline location, by the average velocity across that area.

Volumetric functional ultrasound imaging

To extract the functional response of the rats due to vibrotactile stimulation, changes in blood flow were measured as the intensity of the power Doppler signal. To do this, the data was SVD filtered as described in the previous section, and then 4D power Doppler processing⁸ was performed, where the moving integral of the SVD filtered data, squared, is calculated. Power Doppler processing was performed with an integration window of 200 ms and a temporal step size of 30 ms. To determine if the change in blood flow had regional specificity, the power Doppler intensity through time was plotted in the right cortex, the left cortex, and two control regions, one on the left including the left midbrain and regions of the thalamus and hippocampus and one on the right including the same regions on the right hemisphere. Activation maps were created using the t-statistic at every voxel, calculated using a generalized linear regression model between the power Doppler intensity and the stimulation train. Before this calculation, spatial gaussian smoothing of ten voxels was applied to each frame, with the calculation performed only on voxels within the masked brain region.

Analysis of stimulation response

To analyze the stimulation response, the power Doppler signal through time was plotted, incorporating a baseline normalization step where the average baseline value for the five seconds before the onset of stimulation was subtracted from each stimulation cycle. This accounts for any drift in the data due to changes in microbubble concentration through time. This was followed by a moving average (window length = 15 s) for smoother visual representation. To find the delay between the start of the stimulation and our brain response, the cross correlation was calculated between our signal and the rectangular stimulation signal train in each region, calculating delay as the number of lags between the location of the maximum correlation values. The baseline corrected data with no averaging was used for this analysis. Pearson correlation was used to calculate the connectivity between each brain region⁸³. Then, the square of the Pearson correlation was taken to measure the relationship between the covariance of each of the brain regions. The connectivity matrix was calculated on two time windows, one on the baseline acquisition acquired before the acquisition with stimulation and one during the first cycle of vibrotactile stimulation. The equivalent time for one cycle (70 s) was used to calculate the baseline connectivity. To perform analysis between specific brain regions, the ULM data was used to manually register the data with a volumetric Fischer344 MRI-derived rat atlas^47,54. The ULM data was used to align the data with the structural components of the atlas, particularly the edges of the brain and the larger vessels like the middle cerebral artery. The atlas was scaled separately in each dimension. The young and old rat were registered separately to the same atlas, with the atlas resized for each based on these structures. Sixteen brain regions of interest were selected to remove regions consisting of ventricles and fiber tracks, with the average connectivity calculated between each of the selected regions. The matrices for the baseline and with stimulation cases were plotted on the same scale, with the same regions and ordering.

Combined functional and super resolution images

The functional images show the changes in blood flow, but with a low-resolution representation of the underlying vasculature. A synergistic approach, combining the high temporal resolution of fUS with the high spatial resolution of ULM, was implemented using the same data but different processing methods. The fUS with the ULM images, created on the same dataset, were plotted together, creating a time series with both the functional and vascular information. Since the fUS and ULM images were created on the same dataset, they are registered to one another, and the same vessels showing functional activity will be present in the ULM images. The fUS portion was created by masking the brain using the brain atlas to eliminate noise from outside the brain. The resulting volumetric images were upsampled, creating images with a voxel size of /10. One time point before stimulation and one during stimulation were plotted in 3dslicer⁵³, using the same colorbar and position for each time point. The activation map, calculated using the fUS dataset, was used as a mask to segment the specific vessels within that activated region. Then, structural information can be obtained from those vessels. This allows the creation of activation, connectivity, vessel diameter, and vessel velocity to be calculated on one dataset.

Analysis of vascular response

To analyze the vascular response at high resolutions, the ULM images were used. The activated region was masked using the fUS activation maps, where voxels containing t-statistic values greater than one third of the maximum value were included. Then, the diameter, velocity, and blood flow were extracted from the segmented image. These values were compared both for only the vessels within the activated region and for the brain volume during on and off stimulation periods. To do this, the mean diameter, velocity, and blood flow were calculated for two different cases. In the first case, only frames corresponding to those where stimulation was on were included. For the second, only frames where the stimulation was off were included, essentially creating two different ULM images. Then, the average values between the two cases were compared.

Statistical analysis

To determine if the CBF between on and off periods were significantly different, the stimulation plot from each brain region was split into two vectors, each containing the results from either when stimulation was on or stimulation was off, subdivided into the first four stimulation cycles. Each time point was treated as an individual data point. Then, to determine if the CBF of the on and off periods between the four brain regions were significantly different for each of the five cycles, a two-way ANOVA with a Bonferroni correction was performed. The first independent variable was CBF response in brain region. This included eight groups, consisting of the CBF when stimulation was on and off for each of the four selected brain regions. The other was cycle number, including each of the stimulation cycles with full 30 s on and 40 s off cycles, consisting of the first four stimulation cycles. The diameter, velocity, and blood flow across the brain volume and in the activated region were compared using two tailed Welch’s t-tests. All statistical analyses were performed in MATLAB.

Author contributions

F.S., O.V.F., P.A.D., and G.F.P. conceived the experiments. R.M.J., R.M.D, and H.R.L. conducted the experiments. G.F.P. conceived and developed the beamforming implementation. R.M.J. processed the data. R.M.J., R.M.D, H.R.L., S.M., F.S., O.V.F, and G.F.P. analyzed the results. R.M.J, H.R.L, and H.B. visualized the data. All authors reviewed the manuscript.

Data availability

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-81243-y.