Status and innovations in pre-treatment CT imaging for proton therapy

Abstract

Pre-treatment CT imaging is a topic of growing importance in particle therapy. Improvements in the accuracy of stopping-power prediction are demanded to allow for a dose conformality that is not inferior to state-of-the-art image-guided photon therapy. Although range uncertainty has been kept practically constant over the last decades, recent technological and methodological developments, like the clinical application of dual-energy CT, have been introduced or arise at least on the horizon to improve the accuracy and precision of range prediction. This review gives an overview of the current status, summarizes the innovations in dual-energy CT and its potential impact on the field as well as potential alternative technologies for stopping-power prediction.

Introduction

The range and dose distribution of particles are extremely vulnerable to small deviations in the material-specific stopping power relative to the one of water, the so-called stopping-power ratio (SPR). Compared to photon therapy, the CT conversion in the physical relevant quantity (electron density / SPR) is therefore much more challenging and crucial in particle therapy (PT).¹ Hence, an accurate SPR prediction from CT is essential for particle therapy planning. Furthermore, the definition of CT scan protocols and image reconstruction settings as well as in particular the conversion from CT number (CTN) to SPR are challenging and error-prone processes demanding special attention and quality assurance. This is especially true for PT centers in the preclinical preparation phase, when it is often difficult to allocate resources for such, at first glance, sideline tasks. Moreover, in the last years, the use of dual-energy CT (DECT) for SPR prediction was intensively investigated from different perspectives and the technology found its first clinical applications. However, it is extremely difficult to overlook the different options of DECT application for non-experts and to judge the clinical value from the massive load of scientific publications on this topic.

In this review article, we give an overview of innovations in CT imaging relevant for PT and the state-of-the-art approach for CTN-to-SPR conversion. Secondly, we summarize the status, benefit and potential of DECT for SPR prediction. Finally, we will update on potential alternative technologies for SPR prediction or its refinement.

Current status

CTN-to-SPR conversion via HLUT

X-ray attenuation, measured in CT scans, is based on photon interactions induced by photoelectric effect as well as incoherent and coherent scattering, whereas particle interactions are described by the Bethe equation. Hence, there is no one-to-one physical relationship between tissues’ CT information and its stopping behavior on particles. Since photon and particle interactions share a linear dependency on electron density for human tissues, a stepwise-linear correlation function between CT number and SPR, the so-called Hounsfield look-up table (HLUT), is the standard approach for SPR prediction. In some treatment planning systems, the HLUT is defined between CT number and mass density, which is then internally translated into a material composition and finally SPR.

However, the CTN-to-SPR correlation carries ambiguities since there is also a dependence on effective atomic number (photon interactions) and on mean excitation energy (particle interactions), which cannot easily be included in such a conversion function that depends on a single parameter, the CT number. Hence, the HLUT-based conversion is an intrinsically heuristic approach.

Two different methods exist for HLUT definition: the empirical and the stoichiometric calibration. For the empirical calibration,² both the CTN and SPR of tissue surrogates are measured. Tissue surrogates mimic the X-ray attenuation and, ideally, also the particle stopping properties of human tissues. However, the non-perfect tissue equivalency for both photon and particle interactions is actually the biggest disadvantage of this approach. In contrast, for the stoichiometric approach,^3,4 measured tissue properties (density, elemental composition) of real human tissue samples,^5–7 mostly extracted from corpse, are used to calculate the tissues’ SPR. Only a few datasets of these tissue properties exist - mostly generated by Woodard and White in the 1980s. However, for HLUT definition also the CTN of these tissues is needed, as if they would have been scanned with the institutional CT hardware and the specific scan and reconstruction parameters. Since such a measurement is not possible, the CTN need to be calculated institution-specifically. Hence, tissue surrogates are still required to perform a calibration of such a CTN prediction. To ensure a robust and reliable CTN calculation for tabulated human tissues, a careful selection of calibration materials is still crucial.^8,9 In contrast to the empirical approach, the stoichiometric method only uses tissue surrogates to reproduce the X-ray attenuation of human tissues (not their particle stopping behavior) by preserving similar material parameters such as electron density and effective atomic number (EAN).

The limitations of the HLUT approach can be summarized as follows:

• Limited material differentiation, e.g. different soft tissues cannot be distinguished in CTN but have a different particle stopping power.

• Severe deviations for non-tissue materials (polymers, implant materials), which are usually not tissue-equivalent and are thus not covered by the HLUT.

• Not patient-specific, i.e. the potential intra- and inter-patient variability in CTN-to-SPR conversion is not covered. Inter-patient variations arise from differences in tissue composition,⁷ e.g. different calcium content in bones between adults and children as recently shown¹⁰;

From a survey of 12 mostly European particle centers, the variability in HLUT definition became obvious¹¹: two-third of the centers used a stoichiometric calibration whereas the other third applied an empirical calibration. A specific beam hardening correction for bone, leading to a better CT number constancy for varying scan objects (e.g. patient sizes),¹² was only applied in one-third of the centers. The number of patient-group specific HLUTs per center varied between 1 (5 of 12 centers) and 6. In addition, the number of used line segments for the stepwise linear conversion varied between 2 and 11. Still, due to the dependence on CT scanner and scan protocol, one cannot directly conclude from differences in HLUT definition to differences in SPR and range prediction.

In a recently presented, yet unpublished, study within the European Particle Therapy Network, this inter-center variation in SPR and subsequent range prediction was investigated with a ground-truth phantom scanned at 17 participating PT centers with their individual scan protocols and institution-specific CTN-to-SPR conversions. Alarmingly, the 2σ inter-center variation in range prediction was 2.6 and 2.9% for typical brain and prostate treatment fields, respectively. On SPR level, the inter-center variation was even higher (6.3% in lung, 8.7% in bone surrogates). For about every fifth center, the range deviation from ground truth was exceeding 2%. These results underline the need for standardization of both, CT scan protocols as well as HLUT definition. This will not only improve the individual treatment planning but will also have a positive impact on multicentric clinical trials that are crucial in particle therapy.

Range uncertainty

The before-mentioned limitations in CT-based SPR prediction are the major contribution to the uncertainty in range prediction for PT planning. We are referring here to the nominal range uncertainty based on the planning CT scan - excluding potentially occurring anatomical changes during the course of treatment. The concept of range uncertainties was introduced in practical treatment planning in the 1980s.^13,14 The magnitude of this uncertainty remained practically unchanged since then and can be estimated with a relative term of roughly 3.5% of range plus a 1–2 mm absolute term. Even though, different PT centers use slightly different definitions (e.g. only absolute terms, but different values for different treatment regions), there is a general agreement on the magnitude of range uncertainty.^11,15 In view of the mentioned inter-center variations in range prediction, the overall range uncertainty is at least not overestimated with 3–3.5% for a HLUT-based approach.

General CT improvements for treatment planning

Despite the unchanged uncertainty in SPR and range prediction, several innovations in CT imaging, which are also beneficial for CT applications in radiation oncology, have found their way into clinical application.

Techniques for an automatic exposure control, which adjust the tube current according to the patients’ anatomy, became the clinical standard within the last decade and provide a constant diagnostic image quality in body regions with varying size or severe heterogeneities. Hence, streak artifacts in highly attenuating projection angles can be clearly diminished leading to a more consistent dose calculation and differentiation of anatomical structures. Furthermore, this tube current modulation can contribute to a decrease in CT imaging dose by up to 60%.¹⁶

With the introduction of iterative image reconstruction,¹⁷ it is possible to further decrease the CT imaging dose while maintaining the same image quality (noise), or vice versa, compared to reconstructions using filtered back projection. It was shown that CT dose can be reduced by up to 60% without hampering image quality.¹² Especially, in light of the need of frequent repeated imaging during the course of treatment, such a reduction, enabling doses for control CT scans in the range of 5 mGy (CTDI) or even below, cannot be overrated.

The introduction of advanced iterative metal artifact reduction helped to improve the qualitative image impression with a potential positive impact on target and organ-at-risk delineations.^18,19 However, since these algorithms can also alter the quantitative image information in regions that were initially not affected by metal artifacts,^12,20 their application for dose calculations in particle therapy is not recommended or should at least be carried out with additional inspections and quality control.

Furthermore, algorithms for applications in photon therapy have been developed to estimate the electron density from single-energy CT (SECT) independent from the X-ray tube voltage.²¹ This so-called DirectDensity method distinguishes between bone and water-like contributions and convolutes the effective thickness of bone and water with the respective electron densities in the projection space. The electron-density dataset derived by DirectDensity shares the same limitations for HLUT definition as a SECT scan and is not expected to improve the accuracy in SPR prediction. However, as the DirectDensity approach is independent from the tube voltage used, the tube voltage can be varied, e.g. depending on the patient diameter, and still one single HLUT could be applied.

Dual-energy CT

Idea and motivation

The basic idea of DECT, acquiring two CT scans with different effective X-ray spectra to gain more information about tissue properties, is anything else but new. It was Godfrey Hounsfield, who proposed the concept when he presented the first clinical CT scanner in 1973.²² Only 4 years later, Michael Goitein saw its potential for improved treatment planning in proton therapy.²³ Due to the different relative contributions of photoelectric effect and incoherent/coherent scattering, the two CT scans contain partly complementary attenuation information of tissues. The higher the spectral separation between the two CT scans, the higher is the additional value of the second one. Despite these clear conceptual benefits, it took until 2015 until DECT was first used for clinical treatment planning in PT²⁴ — probably due to the lack of practical hardware and software solutions in the decades before and challenges in data processing, e.g. increased noise level.^25–27

Although the main motivation for DECT in radiation oncology is the improvement of SPR prediction for PT²⁸ with broad consensus in the community,^11,29 several other potential use cases are discussed and investigated: the calculation of relative electron density (RED) for photon treatment planning including brachytherapy^30,31; material classification, e.g. for Monte Carlo applications^32,33 ; improved image quality including metal artifact reduction^24,34 and variable image contrast^24,29 ; target and organ-at-risk delineation^35–39 ; the virtual subtraction of contrast agent information to generate a virtual non-contrast and contrast-enhanced dataset from only one single DECT scan^40–42; functional imaging^43,44 and more.

Technical implementation of DECT acquisition

Within the last decade, various DECT imaging techniques have been technically realized, which can be categorized in single-source and dual-source CT scanners.

Single-source CT scanner

Depending on the type of single-source CT scanner, DECT scans can be acquired

• consecutively (dual-spiral mode),^22,24

• almost simultaneously by fast tube voltage switching,^36,45

• simultaneously using a dual-layer detector,^46,47

• simultaneously using a split-beam filter in scan direction (twin-beam mode).⁴⁸

In some cases, different DECT modes can be realized with the same CT scanner hardware, e.g. dual-spiral and twin-beam.

Dual-spiral DECT benefits from an independent adjustment of tube voltage (commonly 80/140 kVp) as well as current and thus imaging dose. Since motion-induced anatomical changes might occur due to the temporal difference in DECT acquisition (20–90 s), a deformable image registration is performed. For more severe periodic motion, e.g. due to respiration, the dual-spiral technique could in principle be combined with the time-resolved (four-dimensional) CT approach.³⁵ The application of contrast-enhanced DECT scans is limited to a late phase with almost no dynamics starting at least 75 s after injection.

Fast-voltage-switching DECT provides a high spatio-temporal resolution by a fast alternation between low and high tube voltages (commonly 80/140 kV) within 50 µs enabling a projection-based material decomposition. Since the tube current cannot be sufficiently modified for such a rapid voltage change, exposure times for low-energy projections are prolonged (dwell time ratio low:high of 65:35) to allow a higher image dose resulting in comparable image noise. Furthermore, the overall image quality and spectral separation are slightly deteriorated compared to independent acquisitions, because the number of projections per energy and rotation is reduced and a perfectly binary voltage modulation is technically challenging, respectively.

For dual-layer DECT, the energy discrimination is realized solely by the detector arrangement one after the other, leading to different absorption conditions for the same input X-ray spectrum. The rear detector effectively receives a higher photon spectrum. This results in almost perfect temporal correspondence, enabling projection-based reconstruction methods. The spectral separation is considerably smaller compared with other solutions and the measured signal is affected by cross-scatter radiation between detector layers. Thus, no dose modulation is possible between the low- and high-energy dataset.

Twin-beam DECT splits the X-ray spectrum (commonly 120 kVp) into a spatially separated low- and high-energy spectrum by adding an additional filter in a subset of the scanned region (e.g. 0.05 mm gold and 0.6 mm tin filter in a current CT scanner). This technique has an inferior spectral separation and the central part of the X-ray beam cannot be used properly due to the spatial overlap of the two X-ray fields.

Dual-source CT scanner

Dual-source CT scanners consist of two sets of X-ray tube and energy-integrating detector, installed with an angular offset of 9095°.⁴⁹ While simultaneous acquisition with both X-ray detectors, the tube voltage, current and filter systems can be independently adjusted to increase the spectral separation (up to 70 vs 150 kVp with an additional tin filter) and to individually optimize the CT dose. The contamination of each detector signal with cross-scatter radiation is mitigated by corrections during image post-processing. Due to the temporal shift in projection data (quarter rotation time, at least 66 ms), a projection-based material decomposition is very challenging in the presence of patient motion and thus image-based algorithms are preferably applied. Since the CT gantry is limited in space for the two detectors, DECT information is only available in a field of view (FOV) of up to 35 cm. This is the main limitation of dual-source DECT nowadays. Therefore, if a larger DECT FOV of 50 cm is needed, e.g. for treatment planning in the abdominal or pelvic region and for proper inclusion of immobilization devices, a single-source DECT technique seems currently more convenient.

In general, the choice of the optimal DECT acquisition technique and hardware strongly depends on the purpose of the final application (body site, presence of motion etc.) and the relative impact of a multitude of parameters (spectral separation, impact of scattering, current modulation, etc.), and therefore should be subject to exploratory clinical studies to prove the respective benefit in clinical conditions.

DECT-derived virtual monoenergetic images

Depending on the purpose, different post-processing options are available for DECT scans, cf. Figure 1 and Figure 2. A very often used and straightforward approach is a weighted superposition of the two CT numbers derived from DECT (CTN_low, CTN_high), resulting in a so-called virtual monoenergetic CT image (VMI) dataset:

$${\displaystyle CT{N}_{VMI}\left(\alpha \right)=\alpha \cdot CT{N}_{low}+\left(1-\alpha \right)\cdot CT{N}_{high}}$$

With varying the weighting factor α, that can also become negative, the contribution of incohorent scattering (more dominant in the high-energy CT scan) and photoelectric effect (more dominant in the low-energy CT scan) are changed effectively, resulting in different image contrasts. Hence, from the same DECT scan, a VMI dataset of high image contrast (α > 0) for delineation and another VMI dataset with minimized metal artifacts (α < 0) can be reconstructed. For a specific α, the resulting VMI corresponds to the relative electron density.⁵⁰ Alternatively, one can superimpose two VMI datasets to obtain RED as shown for dual-layer DECT.⁵¹

One can also optimize α to minimize remaining beam hardening effects and thereby increasing CT number constancy. Since this is optimal for treatment planning, this approach was chosen for the first clinical implementation of DECT-based VMI datasets for proton treatment planning.²⁴ When using VMI datasets for PT planning, a conventional HLUT conversion from CTN to SPR is still required and all the limitations of the HLUT approach still apply. Nevertheless, compared to a SECT with the same dose, the CT number constancy is increased while the image noise level is reduced.

Direct stopping-power prediction from DECT

According to the Bethe equation, the material-specific SPR directly follows from two physical quantities, namely electron density and mean excitation energy (I-value). The additional material information gathered by DECT can be used to better distinguish between both parameters and almost fully overcome the intrinsic restrictions of the HLUT approach. For this purpose, a variety of algorithms for direct DECT-based SPR prediction have been developed (Figure 1) and thoroughly validated within the last decade, which have recently started to be implemented in PT for treatment planning (Figure 2). In the following, these algorithms are summarized as DirectSPR approaches. “Direct” refers to the direct determination of material parameters (e.g. RED) that can be inserted in the Bethe equation for SPR determination and thereby substantially reduce the impact of heuristic assumptions.

Figure 1. .

Overview of algorithms for direct stopping-power prediction (DirectSPR) based on dual-energy CT.

Figure 2. .

Different levels of clinical application of dual-energy CT (DECT) scans for particle treatment planning. Option A is the combination of virtual monoenergetic CT image and the classical approach using a Hounsfield look-up table (HLUT), whereas in option B the HLUT is adapted with the retrospective application of DirectSPR for a representative patient cohort. Option C is the prospective application of the DirectSPR approach, possessing the most benefits.

Overview of algorithms

The RED is the dominating parameter in SPR prediction and can be directly derived from DECT in a robust and accurate manner^52,53 using either a linear superposition of the two CT scans,^54–56 a stoichiometric calibration,^57,58 a sophisticated modelling of photon attenuation cross-sections⁵⁹ or a projection-based material decomposition in water and compact bone.⁶⁰ Since even with the linear superposition, as the simplest algorithm, a methodological uncertainty within 0.2% can be achieved for a wide range of spectral combinations,⁵³ further algorithmic improvements are not required for PT in regard to other sources of uncertainty.

The EAN is the second relevant quantity for X-ray attenuation and can also be determined by a linear superposition,^61,62 an empirical and stoichiometric DECT calibration^57,58,63 or a sophisticated cross-section model.⁵⁹ Many of these algorithms for RED and EAN calculation^{55,56,61–65} share the same physical principals, differ only in the number of calibration factors and can be mathematically transferred to the equations of Brooks⁵⁴ by applying simple arithmetic conversions.⁶⁶

Various approaches for SPR prediction were developed using RED and EAN as input to solve the Bethe equation. Since the I-value cannot be directly derived from DECT in contrast to RED, an heuristic approximation, similar to a SECT-based HLUT, is needed. The impact of this empirical component on SPR accuracy is however substantially reduced compared to the HLUT approach. In general, either a linear conversion from RED to SPR^60,64 can be performed or EAN is translated to the (logarithmic) I-value using a stepwise-linear⁶⁷ or polynomial function⁵⁷ and then inserted in the Bethe equation.^57,61,65 To better account for tissue mixtures within a voxel, Möhler et al⁶⁸ proposed to determine the relative stopping number (second part in the Bethe equation comprising the I-value) from the photon attenuation cross section (VMI dataset divided by RED) instead of I-value from EAN. Moreover, an empirical SPR parametrization based on the two DECT scans⁶⁹ or the application of machine learning on DECT-derived image datasets,⁷⁰ such as VMI or images of incoherent scattering (comparable to RED) and photoelectric effect (comparable to EAN), were proposed.

Since Monte Carlo dose calculation algorithms require the elemental material composition for each CT voxel, DECT can also be used for an improved material assignment. Based on a pre-defined material database, each voxel can be assigned to the material with the shortest distance in the RED/EAN space.^33,71–73 However, such a material segmentation does not consider intra- and inter-patient variations in elemental tissue composition. In contrast, a RED and EAN parameterization, resulting in a continuous assignment of elemental weight fractions, considers patient variability.³² With a principal component analysis, directly applied on the two DECT scans, image post-processing for RED and EAN determination would not be required anymore.^74–76 Furthermore, an image-⁷⁷ or projection-based^60,78 material decomposition (e.g. water and compact bone) can be performed for material mapping. For all these approaches, the SPR directly follows from the Bethe equation based on the assigned elemental composition.

The comparison between projection- and image-based methods^79–81 as well as a broad variety of image-based algorithms²⁶ did not yield relevant differences in SPR accuracy and precision. However, the calibration of each method, i.e. settings of the CT scan protocol, measurement setup and selection of calibration materials, has an immense influence on the overall performance and robustness.^8,53 Of note, the same is true for SECT-based HLUT approaches, often used as comparison to assess the performance of DECT-based algorithms.

Experimental validation of algorithms

Several studies assessed the accuracy and precision in RED, EAN and SPR prediction in homogeneous tissue surrogates^{26,28,32,33,47,53,56–59,61–65,68,70,72,76,78,82–92} or other polymers^93,94 within simplified phantoms. The superior accuracy of DECT-based methods compared with the HLUT approach could be confirmed in high-precision experimental studies on different elevated complexity levels: (I) various homogeneous^95,96 and heterogeneous^91,92 porcine and bovine tissue samples, which closely represent human tissues, as well as (II) an anthropomorphic head phantom of well-known SPR distribution,⁵⁰ which models usual geometric patient heterogeneities. A DECT-based DirectSPR approach could achieve an SPR accuracy within 0.2% in homogeneous soft tissue samples.⁹⁵ However, a systematic SPR under- or overestimation in small-volume materials at high-density gradients induced by image smoothing during reconstruction was identified, which similarly impacts every CT-based method^50,97 and needs to be considered as CT-related uncertainty.

Due to the high accuracy of DECT-based RED determination and optimized experimental setups for high-precision range measurements, an experimental assessment of I-values in solutions or homogeneous tissues seems feasible.^95,98 Hence, current tissue- or element-specific I-values could be verified or adapted to reduce the uncertainty in SPR prediction.^99,100

First evaluations also demonstrated the feasibility and benefits of quantitative DECT imaging for dose calculation in preclinical studies with mice.^101–103

Clinical relevance for treatment planning

To judge the clinical relevance of the superior SPR accuracy of the DirectSPR method compared with the HLUT approach, dose and range differences were assessed retrospectively on patient datasets.^35,89,90,104 The clinical implementation of DECT for routine proton treatment planning on VMI datasets in 2015,²⁴ still using a HLUT for CTN-to-SPR conversion, provided access to a broad variety of standardized patient DECT scans. Large retrospective patient-cohort analyses of brain-, prostate- or non-small cell lung-tumor cases revealed clinically relevant differences in water-equivalent proton range of 1.2, 1.7 and 2.3% on average, respectively.^35,90 Considering the high SPR accuracy achievable with DirectSPR, the HLUT could be defined/refined based on retrospectively available DirectSPR information derived from DECT scans of patients, which represent a realistic tissue occurrence and can thus provide guidance for HLUT specification (sampling points, course of HLUT). Thereby, systematic range differences can be substantially reduced (within 0.2% on average) resulting in a clear benefit in clinical practice.^12,105 However, to thoroughly account for the intra- and inter-patient tissue variability in SPR,¹⁰ the clinical application of a DirectSPR approach is recommended after careful calibration and comprehensive validation.

Clinical implementation and related challenges

DECT-based SPR prediction approaches are often validated in idealized situations (optimized geometry, central position, high CT dose, large homogeneous materials, calibration for specific beam hardening condition) resulting in SPR accuracies within 1%. However, for a robust and reliable clinical implementation, various influences need to be considered for calibration and validation as well as comprehensive assessment of the relative range uncertainty needs to be performed, which results in clinically realistic estimates of approximately 2%.^28,106

The calibration of each method is crucial and depends on many CT parameters, such as the tube voltage combination, scanner hardware and software as well as scan protocol and image reconstruction settings. Furthermore, different beam hardening conditions induced by varying patient size have a relevant influence on CT number stability and thus calibration factors, in particular if no beam hardening correction for bone is applied during image reconstruction.¹² Since high image noise in DECT datasets can also lead to a systematic SPR deviation,²⁷ iterative image reconstruction and multi-band filtering for noise suppression during image post-processing are highly recommended to obtain an image noise level within 1%, which is sufficient for a precise dose calculation.

Due to the high impact of all these different parameters on SPR accuracy, the calibration should be ideally performed in a standardized and centralized manner or using a commonly agreed scan protocol instead of being center-specific. Hence, CT vendors are encouraged to provide well-calibrated DECT-derived datasets containing material parameters (VMI, RED, EAN, SPR, composition) for their respective DECT technique. These datasets are thus independent from scanner-specific properties and can be universally imported and interpreted by treatment planning systems.

All these influencing variables were included in a recently introduced medical product for patient-specific DECT-based SPR prediction. It comprises a dedicated calibration depending on patient size for a standardized CT scan protocol and a noise-suppression algorithm during image post-processing. The inclusion of these advanced image post-processing algorithms was possible as the SPR calculation was for the first time integrated in CT image reconstruction instead of the treatment planning system. Based on this technology, DirectSPR was clinically implemented for proton treatment planning associated with a reduction of the relative range uncertainty from 3.5% to 1.7–2.0%.¹⁰⁶

Alternative technologies for SPR prediction or refinement

Proton computed tomography

The acquisition of CT datasets using protons instead of X-rays is a promising technology to overcome the uncertainty in the conversion from X-ray attenuation to proton stopping power, because the SPR of human tissues can be directly determined in vivo.^107–110 Despite ongoing research activities over the last decades, this technique is still in an early stage of development, i.e. first prototype proton CT scanners are constructed and experimentally validated, but a clinical system is currently not available.¹¹¹ Current applications are mostly limited to head sizes, since today’s proton therapy facilities equipped with cyclotrons are not capable to deliver proton beams with an energy traversing body regions like abdomen and pelvis. As demonstrated in a recent simulation study,¹¹² proton CT scans might be feasible to be acquired with a total dose smaller than 5 mGy, but further improvements in image reconstruction are crucial to reduce the partly severe ring and streak artifacts. Furthermore, it was shown that DECT-based SPR prediction can achieve a comparable SPR accuracy as proton CT.¹¹²

Proton radiography and range probing

Proton range assessment in patients with proton radiography or so-called range probing is another possibility to mitigate the inherent restriction of a heuristic SPR prediction based on X-ray attenuation. These methods were suggested for a patient-specific HLUT refinement prior to treatment.^113–118 Still, their value for independent validation and quality assurance of the applied range prediction method is barely dispensable. Whereas for proton radiography,¹⁰⁹ a two-dimensional image with information on the water-equivalent thickness of the patient is calculated by deconvoluting the measured signal behind the patient, the range probe approach (also called proton-integrating radiography) directly compares the spot-wise measured depth–dose curve with an expectation derived from dose calculation. Hence, range probing appears to be simpler and easier to implement compared to proton radiography, which has similar limitations and requirements as proton CT. First proof-of-principle implementations and phantom-based validations have been performed.^116,119 Still, neither of the approaches has yet been clinically applied. As additional dose is deposited in the patient (currently 10–20 mGy as determined experimentally for a head geometry¹¹⁸), an approval by ethical and radiation protection authorities is likely to be required.

Prompt-γ-based range measurements

The measurement of prompt γ-rays can provide in vivo information on the actual proton range in patients during treatment without adding any extra dose to the patient.^120–122 Similar to range-probing, a direct in-human validation and potential adaptation of CT-based range prediction for patients seems feasible. For a PGI prototype system, the accuracy and reproducibility was improved, so that the uncertainty in range verification is smaller than the nominal range uncertainty for deep-seated tumors.¹²³ In-human validation of range verification is currently ongoing within a clinical study. Still, the main potential of prompt-γ-based range verification lays in the early and potentially automatic detection of deviations, caused by anatomical changes or patient setup, during fractionated treatments.

MR-only treatment planning

As the application of MRI is substantially increasing in photon therapy, with online-adaptive MR-guided photon therapy being clinical reality nowadays, there is also an increasing need to follow this innovation in PT. This could be by treatment planning solely on MR datasets or, on a longer timescale, even the realization of online adaptive proton therapy guided by MR imaging, so-called MR-guided PT.¹²⁴ However, MR-only dose calculation and also dosimetry in magnetic fields are much more challenging for PT as it already is for photon therapy. Recently developed approaches for MR-based dose calculations generate synthetic CT datasets with heterogeneous tissue representation by an intensity-based assignment of CT numbers, which differentiates soft tissues and bones obtained by segmentation.^125–127 The validation results for small cohorts of prostate- and brain-tumor cases were promising (mean absolute error in CT numbers of 30–80 HU^125,126 or 3 mm error in water equivalent path length¹²⁷). However, further investigations and improvements are required. Depending on the use case, e.g. initial treatment planning or dose re-calculation to assess the adaptation need, different accuracy levels need to be achieved. At the current stage, it remains questionable whether MR-only PT treatment planning can reach a similar level of accuracy and precision as current CT-based approaches, e.g. DirectSPR. For the potential future use of MRI for adaptive treatment purposes, the current research activities are needed to reach and assess clinical applicability.

Photon-counting computed tomography

CT vendors work towards the integration of energy-resolving detectors, allowing for a separation of the X-ray spectrum into multiple energy bins, in future generations of commercial CT scanners. Theoretical and initial experimental studies performed with first photon-counting CT prototypes demonstrated a comparable or slightly better accuracy in SPR and material assignment compared with DECT for the same CT dose.^128–131 The assumption that an increase from 2 to 4 energy bins would automatically lead to an improved accuracy could not be verified in current experiments due to a higher noise level per energy bin and a clear overlap of the bin-specific X-ray spectra. Further developments of spectral denoising methods and projection-based corrections for beam hardening and scattering might be promising to improve the overall performance for more advanced future systems.

Conclusion

In the last decade, the field of pre-treatment CT imaging has seen a lot of research efforts focused especially on PT, resulting in several innovations that already found their way into clinical application or this is on the near horizon (Table 1). The most prominent example is the use of DECT for direct SPR prediction: Substantial and broad evidence for its benefits has been demonstrated on a research level and sustained efforts focused on translating them into real-world clinical implementation have been made. Finally, industry picked up the approach for product development enabling its broad consistent use. DirectSPR has the ability to end the several decades lasting era of heuristic, institution-specific HLUT-based SPR prediction with its intrinsic limitations and thus to substantially reduce range uncertainty. Furthermore, if implemented correctly including inter-center consensus on the algorithm and calibration, inter-center variations in range prediction between centers will be decreased compared to the current situation. On the mid- and long-term scale, several further innovations promise further benefits for PT applications - photon-counting CT being just one example. Still, there is a high need for sustained translational research efforts to adapt technologies and assess their specific benefits for the, in radiology-terms small, application field of particle therapy.

Table 1. .

Overview of techniques for range prediction and its validation

	Applicable for pre-treatment SPR prediction	Applicable for SPR re-assessment during treatment	Technical readiness level**	Current estimated range uncertainty	Advantages	Limitations
SECT-based HLUT	✔	Only with inroom-CT or additional CT acquisition outside of treatment room	9	3–3.5%		see text
DECT-based DirectSPR	✔		8	2%	Consideration of patient variabilities Currently highest accuracy level	No broad clinical experience yet
Photon-counting CT	✔		6	≤2%	Superior detector technology	Systems not mature yet No clinical realistic evaluation
Proton CT	✔	(✔)	4–6	?	Most direct determination	Not yet clinically implemented First systems only for head geometries Severe ring and streak artifacts
MRI-only SPR prediction	(✔) probably inferior to CT-based methods	(✔) for MR-PT integration	6	>2%	Superior tissue classification due to superior soft tissue contrast	Less quantitative image information Difficult differentiation between bone and air Inter-patient variation in bone not considerable
Range probes	✖	✔	6	1 mma	Also direct determination Easy to implement	Only single spots in treatment volume Information mixed with anatomy distal of treatment volume
Prompt-γ-based range assessment	✖	✔	7	1.5 mma	No extra dose or time needed for acquisition Additional recognition of severe anatomical changes	Accumulation of information from multiple spots needed for better counting statistics

DECT, dual-energy CT; HLUT, Hounsfield look-up table; SECT, single-energy CT; SPR, stopping-power ratio.

The TRL definition is according to the Horizon 2020 programme.

^aGlobal verification accuracy of average range shifts within a field only. If subsequently an additional adaptation of the range prediction would be applied, e.g. HLUT refinement, this would bear additional uncertainties - for example an overestimation in one tissue together with an underestimation in another tissue could still exist, even if the integral range shift is minimized on average. Note that both techniques are able to validate an existing range prediction, however cannot replace the (voxelwise) range prediction.