In short, fm-fUS was suitable for rapid classification of active and inactive peaks without requiring any assumption about the shape of the signal for validity.
fm-fUS is an imaging modality that allows detection of brain activity at the mesoscopic scale in freely moving rats, both in cortical and in subcortical areas not accessible to optical methods13. fm-fUS is translatable to other species including primates by adapting the frequency of the ultrasonic probe. A drawback is that fm-fUS, like fMRI, can only measure brain hemodynamics, which are secondary physiological correlates of neural activity, and with a temporal resolution inherently limited by slow changes in blood supply (~1 s). Nevertheless, fUS is compatible with electrophysiological recording at high temporal resolution12,13 to obtain complementary information on neuronal activity.
fm-fUS is currently limited to a single imaging plane at a time. Optical imaging technologies such as intrinsic optical imaging22 or laser speckle imaging23have advantages for applications requiring large FOVs, such as resting-state functional connectivity studies. However, fm-fUS has not yet reached full maturity, and miniature two-dimensional (2D) matrix array transducers or fast-scanning strategies currently under development will enable 3D imaging of the whole brain. Moreover, the weight of the probe can be further reduced to less than 1 g by using transducer technologies similar to those used in intravascular catheters for humans24. This approach would extend fm-fUS to mouse models.
Our work demonstrates that fm-fUS can detect brain activity in cortical or subcortical regions during active tasks. A potential application for fm-fUS lies in BCIs. An important issue with fm-fUS decoding is the delay of 3.5 s between task initiation and detection of the CBV signal. Although this delay appears long, it is comparable to delays observed in steady-state visually evoked potential–based BCIs, which have the highest bit rate among electroencephalogram-based methods (3.4 s; ref. 25). Nevertheless, fm-fUS accuracy and throughput may be increased by analyzing many brain areas simultaneously (multiplexing) and by reducing the processing time.
Considering its spatial and temporal resolutions, penetration depth, real-time processing and functionality, fm-fUS is a promising tool to help understand brain function and may support the development of new types of BCIs.
Methods and any associated references are available in the online version of the paper.
Note: Any Supplementary Information and Source Data files are available in the online version of the paper.
We thank the Ecole Normale Supérieure de Lyon for it’s financial support of the 4th year study project of C. Dussaux. We thank L. Zamfirov and S. Raja for computer-assisted design and technical help with the fm-fUS implant. We also thank the Phenobrain platform of the “Centre de Psychiatrie & Neuroscience” for animal care. This work was supported by a grant from Agence Nationale de la Recherche, Paris.
A.U. and C.D. contributed equally to this work. A.U. and G. Montaldo designed the experiments. C.D. and G. Martel performed behavioral experiments. A.U. designed the head implant and associated surgical procedures. G. Montaldo programmed the ultrasound system. A.U., C.D, C.B. and G. Montaldo performed the experiments. E.M. designed the visual experiments. A.U., C.D. and G. Montaldo analyzed the results. A.U. and G. Montaldo wrote the manuscript, which was edited by all authors.
COMPETING FINANCIAL INTERESTS
The authors declare no competing financial interests.
Reprints and permissions information is available online at http://www.nature.com/reprints/index.html.
1. Hartmann, D.A., Underly, R.G., Watson, A.N. & Shih, A.Y. A murine toolbox for imaging the neurovascular unit. Microcirculation 22, 168–182 (2015). Medline
2. Picciotto, M.R. & Wickman, K. Using knockout and transgenic mice to study neurophysiology and behavior. Physiol. Rev. 78, 1131–1163 (1998). Medline
3. Ferezou, I., Bolea, S. & Petersen, C.C. Visualizing the cortical representation of whisker touch: voltage-sensitive dye imaging in freely moving mice. Neuron 50, 617–629 (2006). Medline
4. Dombeck, D.A., Khabbaz, A.N., Collman, F., Adelman, T.L. & Tank, D.W. Imaging large-scale neural activity with cellular resolution in awake, mobile mice. Neuron 56, 43–57 (2007). Medline
5. Flusberg, B.A. et al. High-speed, miniaturized fluorescence microscopy in freely moving mice. Nat. Methods 5, 935–938 (2008). Medline
6. Sawinski, J. et al. Visually evoked activity in cortical cells imaged in freely moving animals. Proc. Natl. Acad. Sci. USA 106, 19557–19562 (2009). Medline
7. Sullivan, M.R., Nimmerjahn, A., Sarkisov, D.V., Helmchen, F. & Wang, S.S. In vivo calcium imaging of circuit activity in cerebellar cortex. J. Neurophysiol. 94, 1636–1644 (2005). Medline
8. Helmchen, F., Fee, M.S., Tank, D.W. & Denk, W. A miniature head-mounted two-photon microscope. high-resolution brain imaging in freely moving animals. Neuron 31, 903–912 (2001).Medline
9. Miao, P., Lu, H., Liu, Q., Li, Y. & Tong, S. Laser speckle contrast imaging of cerebral blood flow in freely moving animals. J. Biomed. Opt. 16, 090502 (2011). Medline
10. Levy, H., Ringuette, D. & Levi, O. Rapid monitoring of cerebral ischemia dynamics using laser-based optical imaging of blood oxygenation and flow. Biomed. Opt. Express 3, 777–791 (2012). Medline
11. Holzer, M. et al. in Conf. Proc. IEEE Eng. Med. Biol. Soc. 29–32 (IEEE, 2006).
12. Macé, E. et al. Functional ultrasound imaging of the brain. Nat. Methods 8, 662–664 (2011). Medline
13. Urban, A. et al. Chronic assessment of cerebral hemodynamics during rat forepaw electrical stimulation using functional ultrasound imaging. Neuroimage 101, 138–149 (2014). Medline
14. Montaldo, G., Tanter, M., Bercoff, J., Benech, N. & Fink, M. Coherent plane-wave compounding for very high frame rate ultrasonography and transient elastography. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 56, 489–506 (2009). Medline
15. Mace, E. et al. Functional ultrasound imaging of the brain: theory and basic principles. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 60, 492–506 (2013). Medline
16. Liu, J.V. et al. fMRI in the awake marmoset: somatosensory-evoked responses, functional connectivity, and comparison with propofol anesthesia. Neuroimage 78, 186–195 (2013).Medline
17. Rayshubskiy, A. et al. Direct, intraoperative observation of ~0.1 Hz hemodynamic oscillations in awake human cortex: implications for fMRI. Neuroimage 87, 323–331 (2014). Medline
18. Du, C., Volkow, N.D., Koretsky, A.P. & Pan, Y. Low-frequency calcium oscillations accompany deoxyhemoglobin oscillations in rat somatosensory cortex. Proc. Natl. Acad. Sci. USA 111, E4677–E4686 (2014). Medline
19. Lund, R.D., Lund, J.S. & Wise, R.P. The organization of the retinal projection to the dorsal lateral geniculate nucleus in pigmented and albino rats. J. Comp. Neurol. 158, 383–403 (1974).Medline</jrn>
20. Schulz, D. et al. Simultaneous assessment of rodent behavior and neurochemistry using a miniature positron emission tomograph. Nat. Methods 8, 347–352 (2011). Medline
21. Metz, C.E. Basic principles of ROC analysis. Semin. Nucl. Med. 8, 283–298 (1978). Medline
22. Mayhew, J.E. et al. Cerebral vasomotion: a 0.1-Hz oscillation in reflected light imaging of neural activity. Neuroimage 4, 183–193 (1996). Medline
23. Bergonzi, K.M., Bauer, A.Q., Wright, P.W. & Culver, J.P. Mapping functional connectivity using cerebral blood flow in the mouse brain. J. Cereb. Blood Flow Metab. 35, 367–370 (2015).
24. Alves, K.Z. et al. In vivo endoluminal ultrasound biomicroscopic imaging in a mouse model of colorectal cancer. Acad. Radiol. 20, 90–98 (2013). Medline
25. Bakardjian, H., Tanaka, T. & Cichocki, A. Optimization of SSVEP brain responses with application to eight-command Brain-Computer Interface. Neurosci. Lett. 469, 34–38 (2010). Medline
The investigation was performed in accordance with the US National Institutes of Health Guide for Care and Use of Laboratory Animals. The protocol was approved by the Local Animal Ethics Committee of Paris 5 (CEEA 34) and conducted in accordance with Directive 2010/63/EU of the European Parliament and with the ARRIVE guidelines. Seven adult male Sprague-Dawley rats of 8 weeks (Janvier Labs) were used in this study. Animals were kept in a dedicated behavioral room starting 1 week before the beginning of the training for habituation to human and to the environment. Rats were initially kept in group cages in a 12-h dark-light cycle environment at a constant temperature of 23 °C. One week before the beginning of the behavioral training, rats were individually housed with ad libitum access to food and water and were handled daily.
Miniature ultrasound probe.
The ultrasound probe was optimized to achieve minimum size and weight while preserving image quality. Hence, the number of channels was limited to 64 so that we could use a smaller and more flexible cable between the probe and the imaging station, allowing freely moving imaging without providing an excessive burden on the head of the rat. As a consequence, the FOV was restricted to two-thirds of the size brain in a coronal plane (9.6 mm). Moreover, the inter-element distance (pitch) was set to 150 µm to maximize the lateral resolution that is linked to the numerical aperture of the array. The lateral resolution in a point (x, z) of the image can be calculated as:
where n is the maximal numerical aperture of the array, L is the dimension of the array, lambda is the ultrasound wavelength and Dr, Dl are right and left aperture of the array, respectively. A 150-µm pitch enables a numerical aperture of 1.5. Therefore, the lateral resolution of the image is 150 µm in the central part of the FOV and down to 300 µm at the border of the image. Note that the resolution in depth has a constant value of 100 µm as determined by the duration and frequency of ultrasound pulses (Supplementary Note 2).
Four rats were trained to perform a food-reinforced conditioned operant task in a 1.2-m-long straight-alley maze (Fig. 4a) placed 70 cm above the floor over 10 d. During the training, the amount of food ration was limited to 5 g per day to maintain rats at around 80% of their initial weight.
During the task, animals were placed at one extremity of the maze (starting point) and had to reach the other extremity to obtain a reward (AIN-76A Rodent Tablet 45 mg, Sandown Scientific). Then rats had to come back to the starting point and perform the task again to obtain another reward (reinforced round trip, RRT). The reward was placed between six vertical sticks deflecting the whiskers during reward collection (Fig. 4b). All tasks were video monitored, allowing off-line tracking of animal movements. The preconditioning period lasted between 5 d and 1 week, after which the animals were well habituated to the setup and showed no sign of stress throughout an entire training session. When rats were able to successfully perform the task, they received ad libitum access to food and water during 48 h before the surgery to implant the head plate. After 2 d of recovery from surgery, rats performed their tasks with the magnetic probe on the head without the need for habituation and with moderate decline in performance compared to control rats (Supplementary Fig. 7). Food restriction was maintained from day 2 to day 7 as described previously.
Surgical implantation of the head plate.
All components of the magnetic ultrasound probe holder were CAD designed with Sketchup software (Google) and 3D printed in a biocompatible polylactate (Sculpteo). Rats were deeply anesthetized with isoflurane (5% for induction, 2% for maintenance) in 100% O2 and then head fixed in a stereotaxic frame. The eyes of the rats were covered with Vaseline to prevent them from drying out. Prior to skin incision, lidocaine was injected subcutaneously (1 mL per kg body weight), and then a small cranial window of 2 mm × 10 mm was created in the coronal plane (from AP –1.5 mm to – 3.5 mm and L ± 5 mm) over the center of the BF cortex. Then pre-holes were made with a small drill burr without damaging the dura, and the head plate was screwed with six stainless screws (Stoelting) to the bone (Fig. 1a). The brain was covered by low-melting 2% agarose containing ampicillin (100 mg/L, Sigma-Aldrich) to limit infections. Agarose gel ensured proper acoustic coupling required for optimal transmission of the ultrasonic beam in the brain tissue. Rats were placed in a warm cage directly after surgery and monitored periodically until wake up. To reduce pain, we used buprenorphine (i.p. 30 mL per 100 g body weight, Buprecare) immediately after the surgery, and a second injection was performed 12 h later. Finally, the magnetic head shield was placed to protect the brain between each imaging session.
Two rats were used for visual experiments. The surgical implantation of the head plate was performed at 4.5 mm posterior to bregma over the LGN of the thalamus. The rats were returned to their home cage following the surgery and were kept under reduced illumination during visual experiments, which started at day 0 after the complete waking of animals and lasted up to 10 d. Two blue (470 nm, 500 mW) LED strobe lights (OptoLed Lite, Cairn Research) were placed 20 cm from the cage on either side, and the stimulus paradigm consisted of a rest period followed by a period during which the LEDs were blinking at 3.3 Hz (10-ms pulse width, 50% output power) for either 7.3 s (24 flashes) or 2.1 s (7 flashes). The light stimulus frequency was selected to elicit strong binocular visual responses26, and each fm-fUS imaging session lasted up to 1 h, during which period the stimulus was presented many times with a random interstimulus duration ranging from 5 s to 1 min.
Animal tracking in quiet rest and during operant tasks.
We used two HD webcams (Logitech) associated with a free video analysis and modeling tool built on the Open Source Physics Java framework (Tracker 4.87) to track the position of the rat during tasks. In a first set of experiments, we quantified the influence of the fixation of the ultrasound probe on the head of the rat by measuring the time to perform a food-reinforced operant task before surgery or after surgery without or with the probe. We measured the time that was necessary for the rat to move from a starting zone and to reach a reward zone. This delay was measured with 33-ms resolution and allows calculation of the running velocity. We also quantified the residency time in the reward zone as defined by the time to collect the reward. The stimulation pattern, A(t), was considered equal to 1 when the nose was between the sticks, and 0 otherwise. This pattern was used to compute brain activation maps.
Ultrasound sequence and computation of CBV images.
We used a fast sequence for functional imaging that we adapted from the µDoppler sequence described previously27 with the following parameters: five angles (–6°, –3°, 0°, 3°, 6°) averaged three times, 7.5-kHz firing frequency, 500-Hz frame rate, 200 images, 0.4-s acquisition time, and impulse of two cycles at 15 MHz.
This ultrasound sequence triggers emission of a set of 5 × 200 plane waves. We recorded a set of 5 × 200 signals for each channel of the ultrasound array. Each CBV image was calculated using a three-step process: (i) a GPU beamforming step to get a set of 200 ultrasound images a(x, z, t), where x is the lateral position, z the depth and t the time (200 points); (ii) a spatiotemporal filtering step to obtain the blood signal aB(x, z, t) (see “Spatiotemporal filtering” below); and (iii) a computation step of the CBV signal intensity as I(x, z) = ∫aB2(x, z, t)dt. Steps (ii) and (iii) are performed by the workstation.
As a final result, we obtained an image I(x, z) proportional to the CBV signal that we termed ‘CBV image’ for simplification. For detection of hemodynamic variations, we captured CBV images every 0.7 s (1.42-Hz frame rate) during the entire task to obtain a time-dependent signal of CBV in each pixel I(x, z, t).
We developed a specific antivibration filter to minimize the effect of mechanical vibrations during functional imaging in free-movement conditions.
After beamforming of the echoes data, each CBV image was obtained from a set of 200 ultrasound images a(x, z, t) comprising a tissue and a blood component as a(x, z, t) = aT(x, z, t) + aB(x, z, t). We used both temporal and spatial differences between those signals to extract the CBV signal.
The temporal difference was based on the slower velocity of the movement in the tissue compared to blood. Therefore, the Doppler shift of the tissue had a low-frequency signal. The temporal filter was a third-order high-pass Butterworth filter with 50-Hz cutoff frequency. In case of excessive vibration, the temporal filter was not sufficient to separate CBV signal from noise; therefore, we also applied a spatial filter.
Vibration generates a coherent movement in the tissue. In contrast, the blood signal coming from moving red blood cells (RBCs) flows randomly inside the vessel and generates a signal that is uncorrelated between two different voxels. The spatial filter used this difference to dissociate the tissue and RBCs by computing the spatially coherent signal.
We decomposed the spatially coherent part of the signal using a singular value decomposition (SVD) with
In a spatially coherent signal, mi(x, z) is widespread in the whole image, whereas the incoherent blood signal is present in only some pixels. We removed the spatially coherent components where mi(x, z) was widespread. The blood signal became
with Ne = 5 is the number of eliminated singular values.
Functional images of brain activity.
The functional images were computed as the correlation between the CBV images I(x, z, t) and the temporal stimulus pattern A(t).
A z score was calculated using a Fisher’s transform:
A pixel was considered activated when z > 2.5 (P < 0.01 for one-tailed test). Spatial correlation maps displayed all significant pixels superimposed on a reference CBV image.
Resting-state functional connectivity and coherence.
The functional connectivity was established by a seed-based approach3. An ROI of 5 × 5 pixels was averaged and compared to a set of CBV images I(x, z, t) to obtain the CBV signal A(t). The connectivity map was computed by correlating this CBV signal in this ROI with the signal in other pixels.
We used equation (3) with A(t) representing the CBV signal in the ROI.
The spectral coherence Cab(w) between two regions “a” and “b” was computed as
where G is the cross-spectral density defined as
The á ñ indicated an average over different trials, Sa and Sb are the Fourier transform of the CBV signal in regions a and b respectively. We averaged ten recordings of 3 min to obtain the coherence signal.
Real-time data processing.
The ultrasound imaging system is composed of four main devices including the ultrasound electronics (Vantage, Verasonics), the PCI-Express bus, the graphics processing unit (GPU) and the CPU. A pipeline strategy was used to process CBV images in real time in which all parts of the hardware are working simultaneously (Supplementary Fig. 10) according to the following workflow.
1. The ultrasound electronics performed the acquisition of one ultrasound sequence in 0.4 s and stored echo data corresponding to 1,000 plane wave illuminations.
2. The PCI-Express bus transferred data from the electronics to the computer in 0.1 s.
3. The GPU computed the ultrasound images (beamforming) in 0.6 s, and the final output of 200 ultrasound images was transferred to the CPU.
4. The CPU performed the spatiotemporal filtering to extract the hemodynamic information of a single CBV image in 0.2 s.
This strategy allowed parallel processing of images n and n + 1.
An initial delay was needed for the data to enter in the pipeline (1.3 s). Moreover, the frame rate is limited by the slowest step of the pipeline. In our case, the slowest process was the beamforming, which takes 0.6 s. The fUS imaging frame period was fixed to 0.7 s, including 0.1 s to avoid buffer overflow. General-purpose processing on GPUs was of great importance to accelerate data processing in our imaging station. Indeed, the beamforming step used an algorithm that can be easily parallelized to thousands of threads. Moreover, GPU reduced usage of CPU that was devoted to real-time decoding of brain activity (Supplementary Software).
Decoding algorithm of brain activity.
Real-time decoding of brain activity was based on the same algorithm used previously but with an additional processing composed of four steps, including the following.
Elimination of images with high levels of noise due to rapid animal head movement. We computed the total CBV of the image as a set of values In where n is the number of images. Then we computed the mean Ī̄ and the s.d. sI of the set of values In. Image i was eliminated if Ii exceeded more than three times the s.d. (Ii > Ī̄ + 3sI).
Computation of CBV images in the BF cortex. We computed the CBV in the BF region and stored values in a vector sn. The ROI in the BF cortex was defined on the basis of the most-activated voxels for a 1-s short manual stimulus.
Filtering and normalization of the data set. We applied a fourth-order Butterworth high-pass filter with 0.05-Hz cutoff frequency to eliminate very low–frequency oscillations of the CBV. The filter signal was normalized to a mean value of 0, and the s.d. was normalized to 1 as sni = (si – s̄)/ss to compare CBV signal between all rats and sessions.
Threshold detection. When sni was greater than threshold, the brain was considered active and a text message was displayed on the screen (Supplementary Video 1). A threshold of 2 allowed detection of 93% of the events with only 6% of false positives (Fig. 5). As application of the decoding algorithm is a fast processing step (<1 ms), the final processing delay between each frame was about 1.3 s as previously defined (Supplementary Fig. 10).
All data used for statistical evaluations were checked for normality of the distributions. The Kruskal-Wallis test was used to assess the relationship between two variables when the assumption of normality was not respected followed by Dunn’s post hoc test for multiple comparisons. The ANOVA test was used for paired samples followed by Tukey’s post hoc test for multiple comparisons of repeated measures. P < 0.05 (two-tailed) was considered significant.
We chose the sample size for experiments (n = 4 rats) on the basis of literature in the field and national guidelines from the local animal ethics committee. No animals were excluded from analysis, and no randomization or blinding was performed.
Custom code used for GPU processing of fm-fUS images is provided as Supplementary Software.
26. Van Camp, N., Verhoye, M., De Zeeuw, C.I. & Van der Linden, A. Light stimulus frequency dependence of activity in the rat visual system as studied with high-resolution BOLD fMRI. J. Neurophysiol. 95, 3164–3170 (2006). Medline
27. Fox, M.D. & Raichle, M.E. Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nat. Rev. Neurosci. 8, 700–711 (2007). Medline