All images were acquired using the Philips Spectral CT 7500 scanner (Philips Healthcare, Best, The Netherlands) at the Heidelberg University Hospital with a standardized head or body protocol at 120 kV_p. The image acquisition settings and reconstruction parameters for head and body protocols are specified in Supplementary Table 4 , and are based on current state-of-the-art clinical protocols used for particle therapy planning at the Heidelberg Ion Beam Therapy Center (HIT, Germany). Both SECT and DLCT image data are automatically generated from the same raw data set for each acquisition on the Spectral CT 7500 scanner, enabling a direct comparison of the two techniques.

Image reconstruction was performed using the iDose⁴ algorithm at levels 0, 3, and 6 (Philips Healthcare, Best, The Netherlands). The iDose⁴ algorithm uses a hybrid iterative reconstruction technique to reduce image noise, and has levels ranging from 0 to 6, where higher levels correspond to greater noise reduction. In this context, an iDose⁴ level of 0 corresponds to conventional filtered back-projection image reconstruction. For imaging of metallic materials, the Philips orthopedic metal artifact reduction algorithm (O-MAR) (Philips Healthcare, Best, The Netherlands) was also applied.

The SPR prediction accuracy of the Philips Spectral CT 7500 scanner was first investigated using a number of custom cylindrical polymethyl methacrylate (PMMA) phantoms with tissue-equivalent inserts spanning the range of clinically relevant CTNs ( Figure 1 ). Five PMMA phantoms were used to simulate different patient sizes: two one-bore cylinders of height 46.0 cm and radius 5.0 cm (“LCT”, “long cylinder thin”) and 8.0 cm (“LC”, “long cylinder”), two nine-bore cylinders of height 10.0 cm and radius 8.0 cm (“SC”, “short cylinder”) and 16.0 cm (“SCB”, “short cylinder big”), and a roughly human-shaped pelvis. Thirteen tissue-equivalent cylindrical inserts (Gammex Electron Density CT Phantom 467, Gammex-RMI, Middleton, WI, USA) of height 7.0 cm and radius 1.4 cm were used as bore inserts: cortical bone, CB2 50%, CB2 30%, inner bone, muscle, brain, adipose, true water, liver, solid water, breast, bone mineral, and lung. Reference SPR values of these inserts were determined experimentally at HIT by measuring the range shift of a carbon ion beam in a water absorber (Peakfinder Water Column, PTW-Freiburg, Freiburg, Germany). Carbon ions were used for the measurement due to their sharper Bragg peak, reduced lateral scattering, and reduced range straggling compared to protons (22). The inserts were placed in the phantoms in specific configurations to minimize artifacts caused by the high-density bone-equivalent inserts (14).

Furthermore, a selection of materials commonly found in non-tissue implants were scanned for SPR prediction. The metals aluminum and titanium and a carbon/PEEK-titanium composite (icotec ag, Altstätten, Switzerland) were imaged in the LC phantom, along with the special materials PMMA, TECAFORM^® and TECAPEEK^® (Ensinger GmbH, Nufringen, Germany), and Teflon™ (The Chemours Company, Wilmington, DE, USA). In addition, PALACOS^® R + G bone cement (Heraeus, Hanau, Germany) was imaged in a water bath.

The clinical benefits of SPR prediction based on DLCT data were investigated and compared with the currently applied SECT approach by using tissue-equivalent anthropomorphic head (Proton Therapy Dosimetry Head, Model 731-HN) and pelvic (Virtual Human Male Pelvis Phantom, Model 801-P) phantoms (Computerized Imaging Reference Systems, Inc. (CIRS), Norfolk, VA, USA).

Treatment planning optimizations with a dose grid of 1 mm were performed with RayStation Treatment Planning System v10 (RaySearch Laboratories AB, Stockholm, Sweden), using the Monte-Carlo dose engine for proton beams and the pencil beam dose engine for helium and carbon ion beams ( Figure 2 ). The target position for each anthropomorphic phantom was selected such that it was located underneath multiple different tissue-equivalent layers, in order to test the various range prediction methods in heterogenous conditions. For the head phantom, an 8 x 8 x 3 cm³ target volume located at the mid-head was optimized for a physical dose of 1 Gy ( Figure 2A ). For the pelvic phantom, two target types were optimized for a physical dose of 1 Gy: a prostate-like geometry of 52 cm³ ( Figure 2B ) and a 6 x 6 x 6 cm³ target volume ( Figure 2C ).

Treatment planning was initially performed with a conventional clinically-employed SECT scanner (SOMATOM Confidence, Siemens Healthcare GmbH, Erlangen, Germany) with a CT resolution of 0.977 x 0.977 x 1 mm³ (head)/0.977 x 0.977 x 2 mm³ (pelvis). Following plan optimization, forward dose calculations were performed on two additional (image) datasets from the Philips Spectral CT 7500: one using the SECT approach and one using the DLCT approach for SPR prediction.

Dosimetric measurements were acquired at HIT with the OCTAVIUS^® 1000SRS P (PTW, Freiburg, Germany) prototype 2D ionization chamber array detector for proton, helium ion, and carbon ion beam treatment plans, as described in previous works (36). For both phantoms, measurements were performed in the high-dose area and at different positions along the distal edge. For the head phantom, irradiation was performed using the gantry at an angle of 0° with the half-head phantom placed on top of the OCTAVIUS^® detector ( Supplementary Figure 2A ). For the pelvic phantom, irradiation was performed using the horizontal beam line with the half-pelvic phantom placed in front of the OCTAVIUS^® detector ( Supplementary Figure 2B ).

Dose distributions were compared using a 3D gamma analysis (37) for local calculation with a passing criterion of 3%/1.5 mm using a low dose cut-off of 5% of the maximum dose.

CT (image) data acquired using the head and body protocols produced similarly accurate SPR predictions ( Supplementary Tables 5 , 6 ). As such, all reports of SPR prediction accuracy for the remainder of the study are based on CT images acquired using the body protocol, unless otherwise specified.

Similarly, the iDose⁴ level used in image reconstruction was found to have no significant effect on the accuracy and standard deviation of predicted SPR values ( Supplementary Tables 5 , 6 ). Therefore, all results reported for the remainder of the study are based on CT (image) data reconstructed using iDose⁴ level 0 (that is, with minimum additional iterative post-processing), unless otherwise specified.

For tissue surrogates, SPR values predicted using DLCT were consistently closer to reference values than SPR values predicted using SECT in all five phantoms ( Figure 3 ; Tables 1 , 2 ). Pearson’s correlation coefficient (r) and linear regression fitting parameters (α and δ) confirmed higher agreement between measured and DLCT-predicted SPR values compared to SECT-predicted SPR values ( Tables 1 , 2 ). For consistency over all tissue-equivalent inserts, we focus solely on the LCT, LC, SC, and SCB phantoms for the remainder of the study, as not all inserts were imaged in the pelvis phantom.

For the LCT phantom, DLCT-based SPR prediction had a mean overall relative residual of 0.7% (range: [-0.3, 1.4]%) while SECT-based SPR prediction had a mean overall relative residual of 1.8% (range: [-5.7, 5.3]%) ( Figure 3A ). For the LC phantom, DLCT-based SPR prediction had a mean overall relative residual of 0.7% (range: [-1.4, 1.6]%) while SECT-based SPR prediction had a mean overall relative residual of 1.5% (range: [-5.3, 4.7]%) ( Figure 3B ). For the SC phantom, DLCT-based SPR prediction had a mean overall relative residual of 0.6% (range: [-1.1, 1.5]%) while SECT-based SPR prediction had a mean overall relative residual of 1.5% (range: [-5.2, 4.3]%) ( Figure 3C ). Finally, for the SCB phantom, DLCT-based SPR prediction had a mean overall relative residual of 0.7% (range: [-1.2, 1.8]%) while SECT-based SPR prediction had a mean overall relative residual of 1.7% (range: [-4.3, 3.5]%) ( Figure 3D ). Across all four phantoms, the average mean overall relative residual was 0.7% for DLCT-based SPR prediction and 1.6% for SECT-based SPR prediction.

Accuracies of DLCT- and SECT-based SPR predictions across different non-tissue implant materials are listed in Table 3 . DLCT substantially outperformed SECT in predicting SPR values for all non-tissue materials. For the metals aluminium and titanium, SPR prediction accuracy was similar with and without the metal artifact reduction algorithm O-MAR. The HLUT derived in the SECT-based approach is shown together with the eight non-tissue implant materials in Supplementary Figure 1 .

Experimental validation of the DEEDZ-MD method for determining mass density was performed using the tissue-equivalent inserts in the SC phantom, yielding a relative mean deviation of -1.4% compared to the vendor’s provided mass density data (Gammex Electron Density CT Phantom 467, Gammex-RMI, Middleton, WI, USA).

For the SC phantom, the effect of lowering the tube current-time product on SPR prediction accuracy was also investigated. The tube current-time product was lowered from 300 mAs to 250 mAs and 200 mAs with no adverse effect on the SPR prediction accuracy and its standard deviation. Increasing the tube voltage from 120 kV_p to 140 kV_p while using a tube current-time product of 200 mAs resulted in approximately the same volume CT dose index (CTDI_vol) as for the standard clinical protocol. DLCT-based SPR prediction using these CT acquisition settings (140 kV_p/200 mAs) had a mean overall relative residual of 0.6%, which is equal to that of the 120 kV_p/300 mAs protocol.

3D gamma analysis (3%/1.5 mm) using local calculation between SECT- and DLCT-based dose distributions and dosimetric measurements acquired with the OCTAVIUS^® ionization chamber array using the anthropomorphic head phantom revealed substantial agreement between measured and calculated dose distributions ( Tables 4 , 5 ). For all three ion types, DLCT-based dose distributions showed higher 3D gamma passing rates compared to SECT-based dose distributions.

For the head phantom, the 3D gamma passing rates (3%/1.5 mm) were 98.8% (¹H), 97.9% (⁴He), and 97.0% (¹²C) using DLCT for the high-dose area of the target volume. For DLCT, the measured distal position of the 72% (¹H)/55% (⁴He)/53% (¹²C) dose level of the target volume was within 1 mm of the predicted distal position of the respective dose level with 3D gamma passing rates (3%/1.5 mm) of 97.4% (¹H), 84.6% (⁴He), and 86.5% (¹²C) ( Table 4 ).

For the pelvic phantom, the 3D gamma passing rates (3%/1.5 mm) were 99.3% (¹H), 99.4% (⁴He), and 99.5% (¹²C) using DLCT for the high-dose area of the prostate-like geometry and 98.1% (¹H), 99.4% (⁴He), and 93.7% (¹²C) for the high-dose area of the cubic target volume. For DLCT, the measured distal position of the 64% (¹H)/63% (⁴He)/80% (¹²C) dose level of the cubic target volume was within 1 mm of the predicted distal position of the respective dose level with 3D gamma passing rates (3%/1.5 mm) of 99.9% (¹H), 91.1% (⁴He), and 97.6% (¹²C) ( Table 5 ).

In this study, we performed a thorough analysis of the use of the Philips Spectral CT 7500 DLCT system for SPR prediction in particle therapy treatment planning. For this purpose, we experimentally verified DLCT-based SPR prediction accuracy and its impact on dose calculation in particle therapy planning with a Spectral CT 7500 scanner using tissue surrogates and non-tissue implant materials as well as anthropomorphic head and pelvic phantoms. To our knowledge, this is the first study to investigate this second-generation DLCT system for application in particle therapy. Moreover, this study presents the first dosimetric validation of DECT-based dose prediction using anthropomorphic phantoms for helium and carbon ion treatment plans. It is important to investigate DECT-based SPR prediction for helium and carbon ions since the impact of range uncertainty for these ion beams may lead to sizeable biological dose deviation, given the sharp gradients of linear energy transfer (LET) and relative biological effectiveness (RBE) end-of-range (38).

For tissue-equivalent materials, DLCT exhibited greater SPR prediction power, in general, compared to SECT with mean overall relative residuals of 0.6–0.7% for DLCT-based predictions and 1.5–1.8% for SECT-based predictions ( Figure 3 ; Tables 1 , 2 ). Ranges represent the variability introduced by four different phantom geometries. Furthermore, there are individual differences in the tissue-equivalent inserts, as discussed in Faller et al. (22). The larger residuals of SECT-based SPR predictions for certain tissue-equivalent inserts (i.e., bone mineral, brain, and inner bone substitutes) ( Figure 3 ) may in part result from differences between the elemental composition of the tissue surrogate inserts and their real tissue counterparts (39).

In clinical practice, many complicating factors to straight-forward SPR prediction exist, such as the presence of artifact-inducing implants in patients. Therefore, we also validated the use of DLCT for SPR prediction in eight common non-tissue implant materials. DLCT again outperformed SECT, although relative residuals for both approaches were significantly greater than those for tissue-equivalent materials: 1.0–18.4% for DLCT-based predictions, and -6.7%–45.0% for SECT-based predictions ( Table 3 ). To illustrate the importance of SPR prediction for non-tissue implant materials, we consider the example of PALACOS^® R + G bone cement. This common component of artificial joints is made mostly of PMMA and zirconium dioxide. Despite the presence of zirconium dioxide, a high-atomic-number material, the SPR of PALACOS^® R + G is relatively low. The resulting uncertainty in SPR prediction can lead to a particle range deviation of several millimeters when using SECT-based treatment planning. Even if DECT is not implemented for quantitative SPR prediction in clinical practice, spectral image data could still be used to better differentiate between normal tissues and non-tissue implant materials and to help identify properties relating to the stopping power of non-tissue implant materials for contouring and SPR override. For example, using known ED and EAN data sets of commonly used implant materials, comparisons can be performed to quantify relevant physical properties to predict stopping power for unknown materials.

Furthermore, we used the Philips Spectral CT 7500 to experimentally validate the DEEDZ-MD method for determining mass density proposed by Saito (34). Our results showed a mean deviation of -1.4% from the reference value, which is similar to the -1.34% deviation reported by Saito (34). Future work may be dedicated to exploring treatment planning possibilities using mass density data.

We also found a result which suggests that tube current-time product can be lowered by 100 mAs in a simple geometric PMMA phantom without adverse effects on SPR prediction accuracy. Minimizing CT acquisition dose is an important component of CT research, particularly in fields with large pediatric contingents, such as particle therapy.

Finally, we demonstrated the feasibility of using the Philips Spectral CT 7500 to improve particle range prediction by irradiating anthropomorphic head and pelvic phantoms. We showed that dose distributions of DLCT-based treatment plans showed greater agreement with ionization chamber-measured dose distributions than dose distributions of SECT-based treatment plans for proton, helium ion, and carbon ion beams ( Tables 4 , 5 ).

DLCT-based SPR prediction accuracy was previously investigated at HIT using many of the same phantoms and tissue-equivalent inserts as in this study, but with the Philips IQon Spectral CT (22) ( Supplementary Table 2 ). DLCT-based SPR prediction in that study yielded mean overall relative residuals of 0.6–0.9%, compared to the 0.6–0.7% reported here. These results indicate that the SPR prediction accuracy of the Philips Spectral CT 7500 is on par with that of the Philips IQon Spectral CT. However, the Philips Spectral CT 7500 provides numerous other advantages over the Philips IQon Spectral CT ( Supplementary Table 3 ).

A related study using the Philips IQon Spectral CT reported similar SPR prediction accuracy results using mono-energetic images and the same inserts for calibration and evaluation (RMSE=0.6%) (25). Moreover, DLCT-based SPR prediction in this study showed similar accuracy compared to other DECT systems for SPR prediction (12–14, 40). The SPR values of certain non-tissue implant materials used in this work have been previously investigated using dual-source CT (14). The DLCT functionality of the Philips Spectral CT 7500 yielded a similar SPR prediction accuracy as the dual-source CT in this previous study.

As CT technology continues to improve, scanners with DECT capabilities are becoming increasingly available—they are already used for diagnostic purposes at many healthcare facilities. The application of DECT to particle therapy treatment planning could potentially improve patient outcomes. For example, inaccuracies in SPR prediction for pediatric proton therapy planning arising from SECT calibration curves based on adult male tissues may be avoided with DECT (41). Furthermore, SECT-based SPR prediction has been shown to introduce large inter-center variations in SPR, reaching up to 9% between different European institutions (42). Thus, DECT-based SPR prediction might offer more consistent SPR predictions between treatment centers or allow new particle therapy centers to begin treatment with greater confidence in SPR prediction. Moreover, recent work has demonstrated the benefits of even small reductions in range uncertainty to normal tissue complication probability (43), supporting that even small improvements in SPR prediction may be clinically beneficial.

This study demonstrated the feasibility of direct, patient-specific SPR prediction using existing clinical equipment and frameworks. However, in order to use DLCT for SPR prediction beyond a defined research environment, it will be necessary to devise and implement a complete workflow of certified medical products which does not currently exist. To start, SPR DICOM files could be provided as an on-demand spectral result directly from the Philips Spectral CT 7500 scanner instead of needing additional calculation steps using ED and EAN data.

The strengths and limitations of the different DECT or spectral CT acquisition techniques currently available have been described in previous works (7, 44–46) and are summarized in Supplementary Table 1 , with a focus on applications to particle therapy treatment planning. Additionally, Supplementary Table 2 lists selected publications on the different DECT or spectral CT acquisition techniques to provide an overview of the current state of research. The optimal DECT acquisition technique and hardware choice depends on the purpose of the application (e.g., body site, presence of motion) and the relative effect of various parameters (e.g., spectral separation, impact of scattering, tube current modulation) (7), which makes it difficult to give a general recommendation. Imaging with a dual-layer detector enables perfectly temporally and spatially aligned data sets. Moreover, DLCT imaging allows for tube current modulation, a full scan field-of-view coverage, and requires no special mode for DECT acquisition. The dual-layer detector design also enables acquisition of dual-energy data at exactly the same phase of contrast enhancement. Furthermore, the DLCT technique facilitates projection-based material decomposition, allowing for better noise reduction and therefore potentially better material decomposition as compared to image-based methods (47). Nevertheless, spectral separation of DLCT systems is lower than that of source-based DECT systems (44), and spectral signal-to-noise ratio is comparable to that of other commercial DECT systems (48). In addition, DLCT systems carry the risk of cross-scatter occurring between detector layers (45).

In the future, particle CT might have the potential to further improve SPR prediction accuracies and serve as a ground-truth when comparing DECT-based SPR predictions (49). Thus far, precise SPR measurements using proton CT or helium CT are challenging to achieve, and provide a slightly lower SPR prediction accuracy compared to DECT (50).

While the tissue-equivalent materials used in this study are considered valid surrogates for biological tissues, they cannot fully represent the heterogeneity and variable composition of real tissues. Before DLCT-based SPR prediction can be implemented in clinical practice, more studies on biological tissue samples and in vivo systems need to be performed. In addition, measurements in this study were only performed with a male pelvic phantom, introducing a gender data gap. It would be desirable to perform similar measurements with a female pelvic phantom, but at the present time such a phantom does not exist. Furthermore, 4D treatment planning is important for radiotherapy treatments which require motion mitigation and/or consideration during treatment, such as the thorax and the abdominal region. Future work could implement 4D DLCT-based SPR prediction and treatment planning. The large coverage of the Philips Spectral CT 7500 compared to the Philips IQon Spectral CT (80 mm versus 40 mm) ( Supplementary Table 3 ) means that a larger portion of the patient anatomy is covered per gantry rotation of the CT system, leading to potential reduction in motion artifacts. Combining this feature with DLCT-based SPR prediction may enhance 4D CT planning in particle therapy for moving targets.

Finally, other beneficial characteristics of DECT should be investigated for all technical implementations available, including DLCT, to understand the full advantages of the technology. Beyond the computational aspects of DECT-based treatment planning discussed in this work, DECT imaging is expected to provide various opportunities to improve the accuracy of multiple parts of the radiotherapy chain. DECT has been suggested to improve image quality and reduce metal artifacts (51), to improve tumor staging, delineation, and characterization (52, 53), and to contribute to improved normal tissue characterization and personalized treatment through physiological quantification (46). Furthermore, DECT also shows potential for improved dose calculations for treatment modalities other than particle therapy, such as brachytherapy and conventional photon-based teletherapy (51). Finally, as briefly explored in this study and proposed by Albrecht et al. (54), DECT has the potential to reduce total imaging dose. Future work might investigate these varied applications of DECT to radiotherapy.