1.

Introduction

Optoacoustic imaging is an emerging modality in which optical absorption is detected acoustically exploiting the photoacoustic effect, whereby the transient deposition of light energy gives rise to mechanical waves.^1,2 The primary advantage of this approach over pure optical imaging is that resolution does not degrade as fast with increased penetration depth, because the scattering of ultrasound in tissue is orders of magnitude lower than that of light. In particular, MSOT uses multiple excitation wavelengths to obtain separate images of specific chromophores of interest including the oxygenation states of hemoglobin and optical probes such as fluorescent dyes and light absorbing nanoparticles.^3–5

With the ability to detect hemoglobin states, anatomical optical contrast and molecular biomarkers in vivo, MSOT is well suited for cardiovascular imaging.^6,7 Cardiovascular disease (CVD) is an important topic in imaging and studies of atherosclerosis and myocardial infarction are of major interest, both in small animals and clinical settings.⁸ The visualization of targeted dyes and nanoparticles at high resolution could enable unique insights into the mechanisms of CVD and therapeutic responses. For this reason, a significant portion of the literature on optoacoustic imaging describes imaging of the heart, ranging from implementations that require minutes to generate images at single wavelengths,^1,9,10 to high frame-rate systems that capture only microscopic fields of view.⁷ We have recently demonstrated an implementation capable of producing transverse slices through the heart in mice from each laser pulse so that the myocardium can be distinguished from blood in the heart chambers.⁶

The high resolution made possible by MSOT, however, is degraded in cardiac imaging by motion caused by the heartbeat. While multielement ultrasound detection enables real-time optoacoustic imaging, the added spectral dimension needed for identifying specific spectral signatures of chromophores of interest comes with time delays, due to wavelength scanning, resulting in multiwavelength images captured at different stages of the cardiac cycle. The spectral unmixing of single frames from each wavelength can yield unpredictable results because of the arbitrary combination of stages of the cardiac cycle. An alternative approach could acquire several frames for each wavelength and then calculate the average, producing a blurred representation of all stages of the motion, and then apply spectral unmixing. We aim to improve the resulting motion-blurred MSOT images by automatically grouping or clustering images according to their similarities prior to spectral unmixing.

Several established imaging modalities such as positron emission tomography (PET), single photon emission computed tomography (SPECT), computed tomography (CT), and magnetic resonance imaging (MRI) employ triggering, gating, or synchronization schemes that only activate data acquisition during the same stage of the cardiac cycle. This effect would be possible in MSOT imaging too, but would require heart monitoring systems and external triggering of the laser source, causing added complexity and increased acquisition times because the laser would not fire at the maximum possible rate. Instead, MSOT offers a significant advantage in that full frames can be acquired from each laser pulse. Therefore, motion correction of MSOT can also be considered in the context of grouping motion insensitive single frames into sets of images corresponding to similar time points in the cardiac cycle, with the aim of producing images with reduced motion blurring. To this end we considered a $k$-means clustering algorithm to classify multiple images acquired at different wavelengths according to the stage of the cardiac cycle. After subsequent spectral unmixing, we compared the results to MSOT images obtained without clustering.

2.

Materials and Methods

3.

Results

In the single-pulse images we collected, the heart is visible and it is possible to distinguish the myocardium from the blood pool inside the ventricles (Fig. 1). By careful inspection of the images one can visually distinguish frames where the left ventricle is relaxed (diastole) from those where it is contracted (systole), as can be seen in the comparison of two selected frames in Fig. 1. Overall, the images also agree anatomically with the appearance of the myocardium in the corresponding mouse cryosection [Fig. 1(c)].

Fig. 1

Single-pulse optoacoustic images of the heart at 860 nm excitation. LV: left ventricle, RV: right ventricle. (a) Selected image during diastole. (b) Selected image during systole. (c) Corresponding cryosection through the ventricles.

Clustering for each of the five wavelengths resulted in assignments where one cluster contained, on average, 57.8% of the total frames as summarized in Table 1. On inspection we found that the cluster to which the majority of the frames was assigned represented the diastole stage. This result is expected since it is well known that the duration of diastole is longer than systole.

Table 1

Amount of frames assigned to each cluster per wavelength.

Of greater importance, however, are the results after spectral unmixing, i.e., taking all the wavelengths into account. By separately unmixing each of the two clusters, we obtained MSOT images for distinct stages in the cardiac cycle with reduced motion blurring as compared to the images when all frames were averaged (Fig. 2). From the images of deoxyhemoglobin contributions, we can resolve the outline of the left ventricle from the more oxygenated bloodpool. By comparing the unmixed images from the large cluster [Fig. 2(b)] with the smaller cluster [Fig. 2(c)], it is clear that clustering for each wavelength and subsequent spectral unmixing produced distinct images of diastole and systole respectively. The corresponding image without clustering [Fig. 2(a)] is a combination of both and therefore suffers from motion blurring. From a profile through these images [Fig. 2(d)] we observe that the averaged MSOT image over all frames is a mix of diastole and systole and loses some sharpness in the edges of the left ventricle compared to the images after clustering.

Fig. 2

Motion correction for unmixed deoxyhemoglobin contribution. (a) Unmixed image after all 100 images were averaged for each wavelength. (b) Unmixed image after averaging only the larger clusters for each wavelength. (c) Unmixed image after averaging only the smaller clusters for each wavelength. (d) Profile along red line through left ventricle in (a), comparing the images after clustering to the averaged (no clustering) image.

By manually segmenting ROI corresponding to the inside of the left ventricle, we were further able to estimate the areas enclosed by the myocardium (Fig. 3). The area inside the left ventricle is higher in the diastole image [Fig. 3(b): $7.5{\mathrm{mm}}^2$] than in the systole image [Fig. 3(c): $5.2{\mathrm{mm}}^2$]. The averaged (uncorrected) image yields a value between the two [Fig. 3(a): $6.6{\mathrm{mm}}^2$]. These results serve to validate the motion correction clustering approach and indicate potential for the estimation of physiologically relevant cardiac parameters.

Fig. 3

Areas inside left ventricle. (a) All frames averaged (uncorrected). (b) Ventricular diastole. (c) Ventricular systole.

4.

Discussion

We have presented a method for dealing with (almost) periodic motion in real-time MSOT imaging of the heart. Our results in mice show that this method can cluster images according to the stage of the cardiac cycle to which they belong, allowing subsequent spectral processing to be performed separately for each stage, resulting in images with reduced motion blurring compared to when no clustering is performed in the process.

Real-time imaging, or the formation of meaningful images from single laser pulses, is an important feature of MSOT in the context of motion correction. Since the acquisition time for optoacoustic signals resulting from a single pulse is less than 50 μs (for the dimensions of a mouse cross section), motion blurring in individual frames is low. Additionally, each frame is comparable to each other frame in a sequence, allowing simple clustering based on the correlation between single frames. The additional processing of 100 frames each at five wavelengths took approximately 7 s, indicating that the method has potential to be adapted for use during experiments: new frames could be clustered as they arrive, allowing a high-resolution live display.

There are some compromises and limitations involved in resolving motion in MSOT imaging by clustering. Clearly, dividing frames into 2 clusters is the simplest correction possible. Increasing the number of clusters could further reduce motion blurring, but reducing the amount of frames per cluster would simultaneously reduce the signal-to-noise ratio, as there would be less signal averaging. Also, the ability of an algorithm to successfully perform clustering based on motion would be reduced with an increasing amount of clusters, because the distances between the clusters would drop. Generally, in an intuitive approach the number of clusters would be chosen to reflect the amount of motion and the number of components of motion (for example, two clusters for heartbeat and one for respiration) in the imaging scenario in question.

The quality of the MSOT images of the heart presented here, in particular of the unmixed images in Fig. 2 where the entire myocardium of the left ventricle is visible, points to applications in CVD research where studies of myocardial infarction in small animal models can combine measures of hemoglobin concentration and oxygenation with distribution of optical molecular probes to gain multichannel, high-resolution insight into the pathology in vivo. While we have demonstrated the method in the context of cardiac imaging, it can be more generally applied to other anatomical regions and sources of motion, such as respiration. Similarly, there is no reason to restrict this method to small animal imaging: clinical applications of MSOT can equally be affected by motion from respiration, heartbeat, and pulse. We consider motion correction to be an essential part of MSOT imaging, since one of the primary advantages of the modality is its high spatial resolution, which can be degraded in vivo because of motion blurring.