Review of ultrasound image guidance in external beam radiotherapy part II: intra-fraction motion management and novel applications

There are two types of US motion estimation techniques: direct and indirect. Direct methods detect echo motion, whether it corresponds to homogeneous speckle or resolved tissue structure, such as anatomical features. Indirect methods estimate the motion of segmented boundaries (see figure 3).

Zoom In Zoom Out Reset image size

Direct (echo-based) methods: Motion can be estimated without region or boundary segmentation, using (i) US echo fluctuation or (ii) tracking methods, either of which may use the phase or the magnitude of the US echo signal (Eckersley and Bamber 2004). Fluctuation (e.g. Doppler) methods have not been applied to motion estimation in RT and will not be discussed further. Tracking methods, also known as time-domain methods (Hein and O'brien 1993, Eckersley and Bamber 2004), measure displacement as the shift in location needed to re-establish echo correlation (Eckersley and Bamber 2004) and have been used to estimate tissue and phantom motion in RT.

US images possess a grainy structure called speckle (Burckhardt 1978). Speckle is an image pattern that is unique to a region of tissue, and is created by the interference between echoes from US scatterers that are too close to each other to be separately resolved (Bamber 1993, Chen et al 1995). It provides image structure that accurately follows tissue motion, even when no resolved tissue structure is present, so long as the tissue stays in view without excessive deformation or rotation and no changes occur to the US imaging parameters (e.g. US beam direction, frequency, pulse shape or beam shape). Violation of these conditions results in a speckle pattern that changes, or decorrelates, as it moves. When resolved tissue structure is present, its motion may be followed directly, although similar conditions apply as for speckle tracking. Direct methods tend to be based on the application of similarity measures to estimating the motion of regions of US speckle or localised resolved tissue structure (Harris et al 2010); commonly it is a combination of both (figure 3). We therefore refer to all such methods as echo pattern matching.

Echo pattern matching is illustrated in figure 4 for a sector scan. A region of interest (ROI), containing a unique echo pattern, is defined in an US image (acquisition 1). In a subsequent US image (acquisition 2), a pattern matching algorithm searches for the region that best matches the echo pattern in the ROI, identified using a similarity measure such as the normalised cross correlation coefficient (NCC) (e.g. Bonnefous and Pesque 1986), sum of absolute differences (SAD) (e.g. Bohs et al 1993), or sum of squared differences (SSD) (Langeland et al 2003).

Zoom In Zoom Out Reset image size

Indirect (image segmentation-based) methods: Indirect motion estimation methods involve image analysis methods to segment an element of anatomy, e.g. the prostate boundary. Once segmentation has been achieved in each acquisition, a positional measure such as centre-of-mass of the object (or objects) detected may be extracted to provide, over many acquisitions, a time-varying displacement. Such segmentation may, however, not always be possible using US data alone due to a lack of reliable landmarks (Yang and Fei 2012).

Segmentation of a clearly defined object displayed with high echo image contrast (e.g. a fluid filled cavity) can be relatively straightforward (Sarty et al 1998). However, US images of a target such as the prostate are difficult to segment because of relatively low contrast and artefacts such as those due to shadowing, meaning that image processing operators such as edge detectors are inadequate by themselves. More complex methods such as the active contouring (Kass et al 1988) have been developed but require initialisation by user defined curves (Pathak et al 1998). Fully automated 3D segmentation based on atlas registration and texture priors has shown promise when compared with manual segmentation in prostate transrectal US (Yang and Fei 2012). For intra-fraction motion, segmentation would have to be both rapid and fully automated. Fast segmentation has been investigated for image guidance of surgery of the kidney (Ahmad et al 2006) and liver (Angelini et al 2005, Foroughi et al 2006), radio-surgery of the liver (Lee et al 2011) and for ventricular volume estimation in echo-cardiography (Angelini et al 2005, Hansegård et al 2007).

The choice of whether to employ a direct or indirect algorithm (or a combination of both) may depend on the availability and quality of features for reliable segmentation. Direct methods are the most commonly studied and the only ones used to date in a commercial US guided RT system (Lachaine and Falco 2013). A number of interesting direct approaches to motion estimation in long 2D B-mode in vivo US liver sequences were recently presented (De Luca et al 2015).

When implementing a direct echo-based technique it is important to consider a number of factors which can influence the accuracy of the motion estimation (some of these factors are also relevant to indirect (segmentation-based) approaches). Motion estimation accuracy can be affected by factors specific to the tissue (target) of interest, such as depth, speed and type of motion. Others factors are specific to the imaging, data type and algorithmic parameter choices. A full discussion was considered too detailed for this review and the main points have therefore been summarised in table 3.

The first commercially available US based intra-fraction motion monitoring system, the Clarity Autoscan^™ (figure 5) integrates a mechanically-swept 3D US transducer into the treatment planning (computed tomography (CT) suite) and delivery (treatment room) process. A 5 MHz transducer is positioned for transperineal prostate imaging. During treatment, 3D US images are acquired at 2.5 s intervals and registered to a reference US volume using a correlation-based search with reference ROIs centred on pixels within 2 cm of the prostate boundary. Phantom studies demonstrate that  <1.2 mm accuracy and precision of motion estimation can be achieved using Clarity Autoscan^™ (Abramowitz et al 2012, Lachaine and Falco 2013). Intra-fraction motion estimation of the prostate in vivo has yet to be compared to other techniques such as those based on x-ray imaging of fiducial markers (Ng et al 2012).

Zoom In Zoom Out Reset image size

Other US intra-fraction motion estimation techniques being researched are listed in tables 4 and 5, many of which have been confined to phantom investigation (table 4). Some have evaluated the accuracy of intra-fraction motion estimation using ROIs containing resolved features (Abramowitz et al 2012, Schwaab et al 2014) and some with ROIs containing US speckle only (Sawada et al 2004, Hsu et al 2005, Harris et al 2007, O'Shea et al 2014).

Tissues investigated: A handful of in vivo studies (table 5) have shown the feasibility of 2D US-based motion tracking of the diaphragm (Xu and Hamilton 2006), liver (Jacso et al 2009, Rubin et al 2012), prostate (Schlosser et al 2010, Schlosser et al 2011) and lung surface (Rubin et al 2012). Liver motion estimation in 3D was demonstrated by Harris et al (2010). The pancreas can be difficult to visualise with an abdominal US transducer. Omari et al (2015) have assessed the feasibility of using portal vein motion, visualised using Elekta Clarity^™, as a surrogate for pancreatic motion. US has not been used for intra-fraction motion estimation for lung tumours, probably due to the high attenuation of US in lung and the success of optical monitoring of surface tumours. US monitoring of diaphragm motion could however be combined with surface motion to obtain a better estimate of lung tumour motion. Other abdominal tissues, such as bladder and kidney have yet to be studied clinically. The bladder is easily visualised on US and there is evidence to suggest that significant motion of the bladder wall may occur during therapy (McBain et al 2009a). Similarly, the kidneys are easily accessible by US and US-based motion estimation of the kidney may have application to SBRT of adrenal gland metastases (Chawla et al 2009) or in paediatric RT of the upper abdomen (Panandiker et al 2012). It is unlikely that US-based methods for estimating breast tissue motion hold great advantage over surface techniques.

Of the US intra-fraction motion estimation studies that employed in vivo data, comparison with the true motion (and therefore motion estimation accuracy) has either not been fully considered or has been based on manual annotations of features in B-mode images (e.g. De Luca et al 2013). There remains a need for full in vivo investigation of the accuracy and limitations of US-based methods, and therefore a need for improved methods for measuring the true motion. Comparison with other technologies such as kilo-voltage (x-ray) intra-fraction monitoring, which has shown acquisition rates (5–10 Hz) comparable with US and systematic accuracy of  <0.5 mm (Ng et al 2012), or with electromagnetic intra-fraction monitoring devices, e.g. Calypso (Willoughby et al 2006) (temporal resolution of 10 Hz, accuracy  <0.5 mm), with appropriate EM shielding, may be astute.

While intra-fraction motion is inherently 3D, tissues generally exhibit greater motion in the superior-inferior and anterior–posterior motion directions. Left or right motion is often significantly smaller, making it less clinically relevant to RT. This supports the possible use of 2D images, as does low cost real-time imaging and the currently available high frame rate imaging. For example, 2D imaging could be performed with an appropriate left–right (RL) PTV margin to cover the expected magnitude of motion in the RL direction. This does, however, require the transducer to be aligned with the plane of dominant motion, and the development of methods for doing this rapidly and automatically. Direct comparisons of high frame-rate 2D US and slower volume rate 3D US for motion estimation of tissues have yet to be made. As an alternative to direct 3D US motion estimation, Preiswerk et al (2014) have combined 2D motion estimation and MRI-derived population-based motion models to estimate 3D displacement at high rates. The following considers the choice of transducer type, which may be 2D, 3D or biplanar.

2D transducers: A number of studies have investigated the use of 2D transducers for motion tracking in RT (e.g. Schlosser et al 2010, Schlosser et al 2011, Rubin et al 2012, see tables 4 and 5). Compared with 3D US, 2D monitoring decreases the (inter-image) computational load and, with smaller inter-frame displacement (due to higher frame rate), allows a smaller search window, further decreasing computation time. The high frame rate also helps to improve motion estimation by reducing inter-frame decorrelation in direct echo-based methods. At sufficiently high imaging rates (inter-frame elevation displacement  <1 mm), and with a pre-calibration curve to convert decorrelation to distance (Bamber and Bush 1996, Chen et al 1997, Bush et al 2005, Housden et al 2007, Chen et al 2010), decorrelation can estimate out-of-plane motion enabling fast 3D motion estimation with a 2D transducer, albeit with potential elevation direction ambiguity. This method was adopted by Schlosser et al (2010), whereby a drop in the peak NCC value was used to indicate out-of-plane motion and rotation of the prostate during RT.

Biplanar transducers: Biplane imaging offers a high frame rate alternative to full 3D imaging whereby orthogonal imaging planes intersect. (e.g. Hossack et al 2001). Schwaab et al (2014) have investigated its use for RT intra-fraction motion compensation.

3D transducers: The advantages of direct motion estimation with 3D imaging over 2D, if the volume rate is high enough to avoid decorrelation due to rotation and deformation, are (i) the elimination of decorrelation (and consequent loss of tracking accuracy) due to out-of-plane motion, (ii) the unambiguous measurement of all 3 components of the displacement vector, (iii) the possibility to estimate rotations about three axes of rotation and deformation (e.g. Meunier 1998, Saito et al 2009) and (iv) a higher motion estimation precision due to the more unique speckle pattern (Morsy and von Ramm 1998). Unfortunately, mechanically swept 3D US transducers have low volume rate which limits their value for tracking relatively high velocity respiratory-induced motion (Harris et al 2011). Fully real-time (>20 Hz) 3D US imaging (Gunarathne 2013) is still at an early stage of development. Future methods may be based on 2D matrix array transducers, where the 3D volume rate is not limited by mechanically sweeping hardware (e.g. Harris et al 2011, Lachaine and Falco 2013, O'Shea et al 2014) and may reach several thousand Hertz (Byram et al 2010) with improved spatial resolution compared to that provided by equivalent 1D arrays. Bell et al (2012) used a 2D matrix array (operating at 48 Hz) for 3D liver motion estimation in RT, showing that volume rates of 12 Hz are need to accurately estimate cardiac and respiratory-induced liver motion.

Intra-fraction estimation requires the US transducer to remain in contact with the patient throughout treatment. This presents a challenge: finding the transducer placement which allows adequate imaging quality without impacting negatively on the RT workflow.

Dosimetric impact: To date, the potential for a transducer positioned for minimum impact on the dose distribution to degrade target motion tracking has been studied only for skin contact scanning, where changes in the dose delivered to the target may be due to (i) RT beam attenuation when irradiating through the transducer (Bazalova-Carter et al 2015) and radiation scattering by the transducer when it is positioned at the radiation field edge, and (ii) modification of the RT beam angles to avoid the transducer (Wu et al 2006, Zhong et al 2013).

Using a RANDO^® phantom and thermoluminescent detectors Hsu et al (2005) found that a transducer positioned for transabdominal prostate imaging resulted in only a 2.6% change (predominantly in surface dose) to the dose over the volume of the phantom from a single 10MV 10  ×  10 cm² photon beam. Transperineal imaging of the prostate (with the transducer in a stationary holder e.g. Clarity Autoscan) has yet to be studied in this way, but for coplanar treatments should have limited impact along the radiation beam directions. Zhong et al (2013) used planning simulations and a virtual transducer to show that, with the exception of superficial targets, liver stereotactic body RT is feasible with the transducer parallel to the patient axis.

Irradiation through the US transducer, which strongly attenuates the RT beam, is not recommended (Bazalova-Carter et al 2015). To avoid this with a suitable margin, e.g. in VMAT treatment, a simple planning structure could be defined and a strict dose constraint applied to the block during trajectory optimization (similar to metal hip avoidance (Prabhakar et al 2013). For transabdominal prostate imaging, the dose-volume histogram (DVH) for a plan that avoided the transducer was in good agreement with the DVH of a clinically deployed plan (Schlosser et al 2010). The smaller the transducer, the less likely it will interfere with the treatment. As 3D and biplanar transducers are larger than 2D transducers this may be another reason to consider the latter. An alternative approach which uses a radiolucent mechanical scanning assembly containing a single element US transducer eliminates the majority of dense materials typically present in an electronically scanned US transducer array (Schlosser and Solek 2015).

Whilst wireless transducers (e.g. Siemens Acuson Freestyle^TM) hold potential for easier implementation, i.e. no US scanner next to the treatment couch, it is unlikely they will decrease the impact on the treatment plan. For certain treatment sites internal transducers (e.g. endorectal or endovaginal) would give the best spatial resolution and target tracking accuracy, due to short transducer to target distance. The dosimetric consequences of such transducer placement, and opportunities to include them in the planning process, need to be studied.

Robotic transducer positioning: Schlosser et al (2010) developed a telerobotic system to control the transducer position. The system was able to maintain the acquisition of high quality images over time periods relevant to treatment delivery in volunteers. It was also able maintain high quality transabdominal imaging during radiation delivery, robot performance and US target motion estimation of a phantom while a 15 MV beam was delivered.

In another robotic system (Sen et al 2015) the operator and robot share control of the US transducer, which helps to create a consistent body deformation from the force needed to make good contact between the transducer and the body. The system tracks the robot position and contact force used by the operator to obtain a reference US image during simulation, and uses virtual constraints to guide the operator to correctly place the transducer at treatment time. Studies of transducer placement are in progress for various abdominal sites, for both setup (inter-fraction) and delivery (intra-fraction motion) (e.g. Bell et al 2014).

US backscatter characteristics (UBCs), as with other forms of ultrasonic tissue characterisation (TC) (Lerski et al 1981, Linzer and Norton 1982, Waag 1984, Greenleaf 1986, Insana et al 1988, Shung and Thieme 1992), are related to properties of US scatterers, e.g. scatterer size, density, spatial organization and relative acoustic impedance (Lizzi et al 2003). They may be used to detect morphological changes in tissue at the cellular level, and to identify disease (Feleppa et al 2004) for target volume delineation and tissue damage, such as cell death caused by RT (Lee et al 2012), for monitoring treatment response. UBCs can be derived using:

• (1)  
  first-order statistical analyses of the echo amplitudes. For examples see rows 1 to 4 in table 6 (Mountford et al 1973, Lerski et al 1981, Nicholas et al 1986, Bamber 1992, Bamber 1997, Bamber 1998, Hill et al 2004).
• (2)  
  higher-order statistical analyses of the echo amplitudes. See row 5, table 6 (Lizzi et al 1983, Wagner et al 1983, Bamber and Nassiri 1985, Nicholas et al 1986, Valckx and Thijssen 1997).
• (3)  
  spectral analysis of RF echoes (backscatter spectroscopy). Features of averaged spectra may be derived from a regression analysis of the normalized Fourier transform of the RF echo signal, and include the mid-band fit (MBF), US integrated backscatter (UIB), spectral slope (SS) and spectral intercept (SI) (O'Donnell and Miller 1981, Lizzi et al 1983, Feleppa et al 1986), as shown in figure 6.

Zoom In Zoom Out Reset image size

Colour Doppler US (CDUS) imaging, power Doppler US (PDUS) imaging and spectral Doppler US measurement (reviewed by, for example, Evans et al 2000, Eckersley and Bamber 2004) are not truly quantitative techniques, being limited by factors such as Doppler signal angle and depth dependence (Bamber et al 2013). Nevertheless, a large body of work has demonstrated the value of relative characteristics such as those listed in table 5 as non-invasive measures of tumour blood flow and vascularisation (Minasian and Bamber 1982, Wells et al 1997).

Changes in the microvasculature at the capillary level (<100 μm) may occur early in response to treatment (Brown 2002). Intravenously injected gas-filled microbubbles (e.g. Goldberg et al 2001, Stride 2008) of ~1–8 μm diameter can be detected with excellent background tissue echo suppression by using nonlinear microbubble-specific US imaging modes, such as pulse-pulse phase and/or amplitude modulation (e.g. Stride and Saffari 2003, Qin et al 2009). In combination with temporal maximum intensity projection (MIP) imaging, such methods produce US 'angiograms' of tumour vasculature (figure 7) and microvasculature (Shelton et al 2015).

Zoom In Zoom Out Reset image size

Tumour perfusion characteristics can be measured by analysing the kinematics of microbubbles in vessels as small as ~40 μm, using dynamic contrast-enhanced US (DCE-US) (e.g. Piscaglia et al 2012, Saini and Hoyt 2014). Clinical microvascular flow imaging with Doppler is becoming possible without a contrast agent. For example, 'Superb microvascular imaging' (SMI), offered by Toshiba on the Aplio500, is said to use advanced clutter reduction to enhance Doppler SNR. Using ultrafast frame rates (tens of kHz) Tanter and Fink (2014) employed massive averaging to enhance Doppler SNR, which has been used, for example, to visualise microvascular 'activity waves' propagating within the brain (Osmanski et al 2014). Ultrafast imaging also provides more time for US beam steering, to estimate the 2D (and eventually 3D) blood velocity vector, allowing complex flow patterns to be imaged (Hansen et al 2009), a technique with potential for more accurate characterisation of tortuous tumour vasculature. A new method of analysing DCE-US data, to assess spatial dispersion of time-intensity curves, has potential to improve prostate cancer detection (Schalk et al 2015), which may eventually be applicable to improving target delineation in RT. Finally, at appropriate microbubble concentrations, signals can be localised from spatially isolated microbubbles, which facilitates the potential for super-resolution (<10 μm) microvascular imaging (Desailly et al 2013, Viessmann et al 2013, Christensen-Jeffries et al 2014).

Microbubbles conjugated to targeting molecules such as peptides or antibodies allow USMI (e.g. Wen et al 2014) of circulating molecules or those associated with endothelium such as vascular endothelial growth factor receptor type 2 (VEGFR-2) (Korpanty et al 2007, Rychak et al 2007) and α_v-integrins (Kiessling et al 2012). Liquid phase-change nanodroplets may allow extravascular molecular targeting but they are at an early stage of development (Wilson et al 2013). Another future option may be protein-shelled gas nanovesicles (GVs) extracted from microorganisms such as bacteria which produce GVs to control their buoyancy. These GVs provide stable US contrast, detectable in vivo with various useful properties as molecular reporters (Shapiro et al 2014). Most USMI work at present is pre-clinical, e.g. for assessment of tumour angiogenisis and response to anti-angiogenic therapies, although initial clinical studies are now underway in prostate (NIH 2015a, NIH 2015b, Kaneko and Willmann 2012).

US elastography aims to display contrast for, or quantities related to, the shear elastic moduli of tissue (Bamber et al 2013). The available techniques (e.g. Parker et al 2011, Bamber et al 2013, Doyley and Parker 2014, Shiina et al 2015) use methods described in section 2.2.1, such as similarity search or Doppler, to measure the temporal and spatial dependence of tissue displacement associated with the shear deformation of tissue, which can be used to infer, or quantify, tissue elasticity. Qualitative measures of elasticity strain and displacement are provided by strain elastography (SE) and acoustic radiation force impulse (ARFI) imaging, respectively. Methods giving quantitative elasticity measures (e.g. shear-wave speed, c_s, related to Young's Modulus, E, by E  =  $3\rho c_{\text{s}}^{2}$, where ρ is the mass density) include transient, vibrational and shear-wave elastography (TE, VE and SWE, respectively). For a detailed explanation of these techniques please refer to (Bamber et al 2013, and Shiina et al 2015).

For RT, elastography may improve the ability to visualise disease for target delineation (see section 3.2). Furthermore, as elastography is sensitive to tumour cell necrosis (Li et al 2014), oedema (Berry et al 2008) and fibrosis (Bush et al 2005), and as tissue elasticity is affected by RT (Yarnold and Brotons 2010), it is plausible that elastography may be used to monitor tumour response and normal tissue toxicity. Finally, as tissue viscoelasticity has been shown to be associated with both fibrosis (Cosgrove et al 2013, Ferraioli et al 2015) and microvessel density (Jugé et al 2012, Jamin et al 2015), and fibrosis is linked with induction of angiogenesis via hypoxia inducible factors (HIF) (Ruthenborg et al 2014), it is conceivable that elastography may help define the BTV. This hypothesis has yet to be tested.

Photoacoustic imaging (PAI) (Wang 2009) uses pulsed light to induce a transient (<10 ns) temperature and pressure rise wherever the light is absorbed, typically in haemoglobin, melanin and, to some extent, fat. This generates acoustic emissions detectable by an US transducer, providing real-time imaging of the optical absorption properties of tissue. Although imaging depths of up to 7 cm have been reported (Zackrisson et al 2014), the reliability of the image detail at such depths depends on whether image clutter is generated by strong photoacoustic sources outside the imaged region, and methods of distinguishing clutter from reliable detail are under development (Jaeger et al 2012, Jaeger et al 2013, Alles et al 2014). In different forms, PAI spectroscopy (Wang 2009), allows (a) molecular identification of endogenous chromophores or exogenous contrast in the form of dyes or nanoparticles which may be molecularly directed to extravascular targets (Wilson et al 2013, Zackrisson et al 2014); or (b) a PAI analogue of US backscatter spectroscopy (section 3.1.1) which permits absorber (e.g. microvessel) size, spacing and spatial organisation to be measured (Gertsch et al 2010, Preisser et al 2013, Xu et al 2014, Li et al 2015). Most PAI studies have been preclinical or have used tomographic systems (e.g. for imaging the breast), although translation to the clinic using hand held US transducers is underway (e.g. Wang et al 2011, Jaeger et al 2012, Alles et al 2013, Montilla et al 2013).

US backscatter spectroscopy (section 3.1.1(3)) has been investigated at high frequencies (>20 MHz) for monitoring radiation-induced apoptosis in vitro and in small animal models (Czarnota et al 1999, Vlad et al 2009), with validation against histopathology (Czarnota et al 2007). Its clinical feasibility to monitor response of deep-seated tumours has yet to be proven, although reason to be optimistic is provided by the ability of the method at low frequencies to characterise microscopic structure in phantoms (e.g. Insana et al 1990) and, for example, to identify glomeruli as one of the dominant scattering structures in the renal cortex (Insana et al 1991).

PDUS and DCE-US have been investigated pre-clinically to quantify tumour response to RT, although some groups found high positive correlation of Doppler characteristics with histologically assessed microvessel density pre- and post-RT (Donnelly et al 2001, Kim et al 2006), whereas others did not (Fleischer et al 1999, Denis et al 2002, Hwang et al 2010). Hwang et al (2010) and Krix et al (2003), however, did report a strong correlation of DCE-US I_p with MVD. Finally, the PDUS vascularity index has been noted to decrease in tumours treated with US and microbubbles (see section 3.3.3), alone or with radiation (Czarnota et al 2012, Tran et al 2012).

Clinically, a significant reduction in pulsatility and resistance indices of the intranodal vessels of metastatic lymph nodes (see table 5) 8 weeks post-RT was demonstrated (Ahuja et al 1999). Pirhonen et al (1995) found RT caused a significant decrease in cervical tumour vascularity during treatment, which was associated with disease free survival. Huang et al (2013) demonstrated a 50% reduction in 3D PDUS vascularity index, 5 weeks from start of RT, dropping to 100% reduction at 3 months, in women receiving RT or concurrent chemoradiotherapy for cervical carcinoma. Krix et al (2005) observed a decrease (~20%, larger than measured using contrast-enhanced CT) in tumour arterial phase DCE-US I_p 2 months after single fraction stereotactic body RT to liver metastases. Following proton therapy of hepatocellular carcinoma, contrast enhanced CDUS tumour CPD initially increased in 50% of patients and then, 9 months post-RT, decreased across all patients (Niizawa et al (2005).

Further work and greater standardisation of techniques are required to understand how Doppler US and DCE-US characteristics relate to tumour vasculature and perfusion. An inherent difficulty in histological evaluation of tumour response both clinically and pre-clinically is establishing accurate spatial registration between histological sections, dose distribution and the US imaging plane. Three-dimensional US, provides more accurate image registration compared to 2D (Hwang et al 2010, Huang et al 2013). Pre-clinically, integration of functional 3DUS into small animal dedicated irradiation devices may bring about increased accuracy in correlative studies of this kind (Verhaegen et al 2011).

Although US has long been studied as a technique for inducing hyperthermia (Hand and Haar 1981), and hyperthermia and RT are known to act synergistically (e.g. Miller et al 1977, Holt 1980), the potential for synergism between therapeutic US methods for example high intensity focused ultrasound (HIFU) (Baker et al 2001, Tachibana 2004, Koonce et al 2015, Wood and Sehgal 2015) and RT has only just begun to be investigated, and only in a preclinical context. The induction of endothelial cell apoptosis by exposure to US in the presence of microbubbles is proposed as a possible mechanism for observed radioenhancement (Al-Mahrouki et al 2012), and a ten-fold increase in tumour cell-kill was observed using a combined treatment relative to microbubbles or radiation alone (Czarnota et al 2012, Tran et al 2012). Furthermore, US contrast-mediated vascular permeation and drug delivery (Rapoport et al 2007) holds potential for the localised delivery of radiosensitisers such as Paclitaxel (Liebmann et al 1994) or epidermal growth factor receptor inhibitors (Sartor 2004).

Tumour hypoxia has been shown to be associated with a decrease in vascularity (West et al 2001) motivating research into Doppler and DCE-US characteristics as markers of hypoxia, validated against measures of oxygen partial pressure (pO₂) using polarography. Clinically, Scholbach et al (2005) and Gagel et al (2007) reported moderate correlation with a tissue perfusion index (see table 6) and polarographic measurements in metastatic lymph nodes, respectively. Pre-clinically, weak (Ohlerth et al 2010) or no correlation (Elie et al 2007) with Doppler US characteristics was found. Poor results may have been a manifestation of problems with various factors in these types of experiment, viz., reliability of polarography in vivo and spatial sampling error of both polarography and US.

Using photoacoustography, Sun et al (2012), measured increases in tumour volume, and decreases in blood concentration and oxygen saturation, in tumour models irradiated with 30 Gy, comparing measurements taken before and 10 d post-irradiation. Similar findings, in response to US and microbubble enhanced RT, are reported by Briggs et al (2014).

Limited clinical evidence to support the use of elastography to measure response to RT exists. Mabuchi et al (2015) using real-time SE, found that in patients with complete response, tumour stiffness decreased to levels similar to normal cervix. Rafaelsen et al (2013) detected a significant decrease in shear wave speed in rectal tumours and mesorectal fat, between baseline measurements prior to the start of chemoradiation therapy (3.13 ms⁻¹) and two (2.17 ms⁻¹) and six (2.11 ms⁻¹) weeks after treatment began. A softening of metastatic cervical lymph nodes post-chemoradiation therapy was demonstrated using SE (Furukawa and Furukawa 2010).

The potential for USMI to measure response to RT has been demonstrated pre-clinically: microbubbles targeted to ICAM-1 (a marker of inflammation) and α_vβ₃-integrin (a marker of angiogenesis) allowed 30 MHz US to image early vascular response of xenograft prostate tumours irradiated with carbon ions, suggesting an increased expression of ICAM-1 and α_vβ₃-integrin in response to RT (Palmowski et al 2009). Gold nanorods with resonance peaks at 700 nm and 900 nm were functionalised with antibodies targeted to HER-2 and EGFR transmembrane receptors (Shah et al 2014) demonstrating potential for contrast photoacoustic stratification of tumours prior to therapy.

Xerostomia is a common complication of dose to the salivary glands during head and neck RT (Eisbruch et al 2001). On B-mode US, parotid glands become more heterogeneous and hypoechoic post-RT (Ying et al 2007, Cheng et al 2011). Significant differences in UBC (Yang et al 2012, Imanimoghaddam et al 2012), spectral Doppler characteristics (Ying et al 2007, Wu et al 2011), and shear wave speed (Badea et al 2013) between irradiated and non-irradiated tissue have been observed in parotid or submandibular gland, although Imanimoghaddam et al (2012) saw no difference in spectral Doppler characteristics between baseline and 6 to 7 week post-RT. No study has examined the spatial distribution of US characteristics; this may be of particular interest for trials of parotid sparing (Chao et al 2001) which aim to spare the region of the parotid containing stem cells.

In patients receiving head and neck RT, Leung et al (2002) detected significant differences (~25%) in subjectively estimated relative Young's modulus between regions of the neck receiving boost irradiation and those that did not. Zheng et al (2000) showed that skin thickness, MBF and SI of neck tissue significantly differed between patients following head and neck RT and healthy volunteers.

Women undergoing adjuvant RT for breast cancer may experience breast swelling (oedema), an acute effect, and breast hardening (thought to be associated with fibrosis), a late effect (Yarnold et al 2005). Yoshida et al (2012) measured skin thickness, subcutaneous tissue lateral correlation (see table 5) and MBF, in the irradiated and un-irradiated breasts of breast cancer patients post-RT (Liu et al 2010). Patients with acute (<6 months) radiation toxicity had significant differences between breasts in all three characteristics. Those with late (>6 months) Radiation Therapy Oncology Group (RTOG) grade 1 or 2 skin toxicity (RTOG 2015) had greater skin or subcutaneous tissue US characteristics compared to patients with RTOG grade 0 skin toxicity.

B-mode US has been used to measure an increase in skin thickness (p  <  0.001) and relative US backscatter amplitude of subcutaneous tissue (p  ⩽  0.05) post surgery and RT (Adriaenssens et al 2012). SE was used to measure a decrease (p  <  0.05) in subcutaneous tissue stiffness in patients approximately 10 weeks post RT compared to pre-surgery (Adriaenssens et al 2012) and an increase in stiffness in the irradiated breasts of women with late toxicity compared to the unirradiated breast (Bush et al 2005).

The risk of heart disease subsequent to RT increases with mean heart dose (Darby et al 2013). Erven et al (2011) used strain rate imaging to show that left-sided breast cancer patients had a significant decrease (p  <  0.001) in myocardial strain post RT (17.6  ±  1.5%) and 2 months follow-up (17.4  ±  2.3%) compared to pre-RT (19.5  ±  2.1%). No decrease in strain rate could be observed in segments that received less than 3 Gy but this may be due to the sensitivity of the technique rather than a lack of effect (Darby et al 2013).

In patients receiving RT for head and neck cancer, spectral Doppler PI, RI, and PSV of the inferior thyroid artery underwent significant changes by fraction 10 of RT, and mean relative US backscatter amplitude and heterogeneity of the thyroid gland decreased (Bakhshandeh et al 2012). Krix et al (2005) showed that post-RT arterial phase DCE-US peak intensity increased in liver metastases suggesting radiation-induced hypervascularisation and support for the hypothesis that radiation-induced liver disease is a result of liver venous obstruction.