The Role of Low-Energy Virtual Monoenergetic Images Derived From Dual-Energy Computed Tomography in Tumor Segmentation for Head and Neck Radiation Therapy: A Contouring Study

Abstract

Purpose

Accurate segmentation of head and neck (HN) tumors on computed tomography (CT) is made more challenging by limited soft-tissue contrast. Dual-energy CT (DECT), through virtual monoenergetic images (VMIs), offers improved tumor visualization, yet its impact on tumor segmentation remains underexplored. This study evaluates the effect of low-energy VMIs on the delineation of HN tumors within a multimodal imaging framework.

Methods and Materials

VMIs were created at energies from 40 to 60 keV and were evaluated for optimal tumor visualization using contrast-to-noise ratio (CNR) metrics and qualitative physician preference assessments. The 50 keV VMI, identified as the optimal compromise between contrast and image quality, was incorporated into a contouring study involving 16 patients with HN cancers. Six physicians adjusted positron emission tomography–derived gross tumor volumes (GTVs) sequentially, incorporating the following images: 120 kVp-equivalent DECT (standard planning CT), 50 keV VMI CT, and magnetic resonance image (MRI). Inter- and intra-physician variability were assessed using Dice coefficients, Hausdorff distance, mean distance to agreement, and contour volume. Qualitative metrics, including tumor conspicuity, edge sharpness, and contouring confidence, were measured.

Results

The 50 keV VMI demonstrated significantly higher CNR compared with the 120 kVp (CNR_50 keV = 1.99 vs CNR_120 kVp = 0.54, P = 6.5 × 10⁻⁷) and was favored by physicians for tumor delineation. The addition of 50 keV VMIs to the 120 kVp image increased GTVs with a mean volume increase of 2.0 cm³. The VMI-based contours were more consistent with findings on MRI. Intra-physician variability was less between MRI and 50 keV VMI contours than between MRI and 120 kVp contours, whereas inter-physician variability remained unchanged or increased slightly with the addition of each image set. Anatomic site influenced VMI utility, with less impact observed at the base of the skull.

Conclusions

Low-energy DECT VMIs, particularly at 50 keV, enhance HN tumor visualization and lead to larger GTVs when compared with 120 kVp images alone, demonstrating consistency with MRI tumor delineation. These findings support the integration of VMIs into multimodal radiation therapy planning workflows to optimize target delineation accuracy.

Introduction

Accurate segmentation of head and neck (HN) cancers on computed tomography (CT) remains a substantial challenge, primarily because of the limited soft-tissue contrast inherent to conventional CT imaging. Although the administration of iodine-based contrast agents enhances tumor visibility, dual-energy CT (DECT) offers additional potential through the generation of virtual monoenergetic images (VMIs).1, 2, 3, 4, 5, 6, 7, 8 It has been hypothesized that the improved contrast between the gross tumor volume (GTV) and surrounding tissues afforded by DECT could lead to more precise target delineation.^9,10 However, there are few studies investigating the impact of DECT on tumor segmentation accuracy.^11,12

To date, 1 study has assessed HN contours created with DECT imaging. Ng et al¹² evaluated inter-observer variability in GTV contouring for patients with recurrent HN cancers, comparing contours derived from 60 keV and 140 keV VMIs. Although overall inter-observer agreement was modest across all modalities (mean Dice Similarity Coefficients ranging from 0.45 to 0.58), the study reported reduced variability in contours generated from the 60 keV VMIs. Their study was constrained by the lack of comparison with conventional single-energy CT (SECT) and does not address intra-observer variability.

In clinical practice, low-energy VMIs are seldom used in isolation for HN radiation therapy planning. Rather, they are typically integrated with other imaging modalities such as conventional 120 kVp CT, magnetic resonance imaging (MRI), and positron emission tomography (PET). A 120 kVp-equivalent image is often incorporated alongside low-energy VMIs during the treatment planning process because low-energy VMIs are more susceptible to increased image noise and artifacts, particularly those arising from patient motion and metal implants. Use of the 120 kVp-equivalent image also enables application of the standard 120 kVp CT number-to-density calibration curve for dose calculation. Furthermore, many autocontouring algorithms are trained on 120 kVp-equivalent images, and additional work is required for low-energy VMIs. MRI has been increasingly incorporated into HN cancer tumor delineation, particularly for nasopharynx, paranasal sinus, tongue, and skull base tumors, due to the improved soft-tissue contrast, which can theoretically improve tumor delineation. MRI simulation and high-quality HN MRI are also increasingly available at more centers. Several studies of the impact of MRI on tumor segmentation have been performed and generally show relatively high inter-observer variability, suggesting the need for greater experience and training with this modality.13, 14, 15 Low-energy VMIs are another method for improving soft-tissue contrast and could serve a complementary role to MRI or even potentially substitute for it in environments where MRI is not readily available. Low-energy VMIs have the distinct advantage of not requiring image registration to the treatment planning CT, in contrast to MRI or PET/CT. Changes in position with MRI or PET/CT can make use of these studies challenging due to imperfect registration to the planning CT. Understanding the utility of DECT VMIs within a multimodal imaging framework requires a comprehensive evaluation incorporating both inter-observer and intra-observer variability, as well as quantitative measurements of contrast-to-noise ratio (CNR).

The aim of this study is to quantify the effect of low-energy VMIs on the delineation of HN tumors through a novel contouring study. First, the optimal VMI energy level for HN tumor visualization is identified using a combination of quantitative CNR metrics and qualitative physician preference assessments. This optimal VMI energy level is then incorporated into a multimodality contouring study to evaluate its impact on inter- and intra-physician variability in HN tumor segmentation.

Methods and Materials

After institutional review board approval, a retrospective study was performed for patients who were treated for HN cancer at our institution using radiation therapy between October 2018 and March 2021. All patients had histologically proven HN cancer. Eligible patients had MRI and CT scans acquired within 30 days and in similar anatomic positions, excluding cases with marked neck flexion differences. All patients were simulated on the Siemens SOMATOM Definition Edge CT scanner for radiation therapy planning purposes. To minimize motion, patients were immobilized with custom head rests and thermoplastic masks. Patients were instructed not to swallow during image acquisition. Each patient received intravenous iodine contrast medium, OMNIPAQUE, with a 60-second delay before CT simulation.

CT images were acquired using a split-filter technique with an effective mAs of 1350. The automatic tube current modulation was not used, and the volume CT Dose Index (CTDI_vol) was 23.2 mGy. Images were acquired with a pitch of 0.45, a rotation time of 1 second per rotation, and reconstructed at a slice thickness of 2 mm and a field of view of 50 cm. For each CT acquisition, 2 image sets were created: a composed 120 kVp-equivalent image and a VMI. The 120 kVp-equivalent image was reconstructed and used as the primary treatment planning CT image for dose calculations. The VMI was created using Siemens Syngo.via software (VB30). The VMI was fused to the 120 kVp image to aid in contouring.

The MRI and PET image sets were registered to the 120 kVp-equivalent image set. Image registrations were performed in the GE Healthcare MIM software using a local rigid registration focused on the site of disease. Registrations were reviewed by a board-certified medical physicist with over 12 years of experience. No registrations were required to fuse the 50 keV VMI to the 120 kVp-equivalent image sets because they were derived from the same image acquisition.

Optimization of monoenergetic energy

Contrast-to-noise assessment

To determine the optimal energy of the VMIs, the average contrast and CNRs were calculated for 20 patients using regions of interest (ROIs) placed in the GTV and the geniohyoid muscle. VMIs were generated at energies of 40 keV and 50 keV and compared with the composed, virtual 120 kVp-equivalent images. Contrast and the CNR were calculated as

$$\text{Contrast}=H_{GTV}-H_{muscle}$$ (1)

$$\text{CNR}=\frac{H_{GTV}-H_{muscle}}{{\sigma }_{muscle}}$$ (2)

where $H_{GTV}$ and $H_{muscle}$ are the average CT numbers in Hounsfield Units (HU) in ROIs placed in the GTV and the geniohyoid muscle, respectively, and ${\sigma }_{muscle}$ is the SD of the CT numbers in the geniohyoid ROI. P values were calculated using a single-tailed paired t-test.

Physician preference

It is well established that imaging features other than CNR can affect tissue segmentation, such as perceived texture, image noise, and image artifacts.^16,17 Therefore, an internal, qualitative study was designed to gauge physician preference. Based on the CNR results, a series of low-energy VMIs was selected for the qualitative study to determine optimal VMI energy. The study included 10 HN patients. Monoenergetic reconstructions were created at energy levels of 40 keV, 45 keV, 50 keV, 55 keV, and 60 keV. Four physicians were asked to rank the images based on their perceived ability to delineate tumors. The average rank and SD for each energy level image were calculated.

Contouring study

Sixteen patients (11 male and 5 female) were selected for this study. The median age (range) was 61 years (27-85 years). Although immobilization was not used for diagnostic PET and MRI, patient selection was limited to patients with similar anatomic setups in PET, MRI, and CT. Twelve tumors were located in the base of tongue (BOT) and tonsil, 3 tumors were located in the nasopharynx, and 1 tumor in the thyroid. Six physicians (the 4 physicians from the “physician preference” study and 2 additional physicians) were provided with contours created by a board-certified medical physicist. Contours were derived from the PET scan using an auto-SUV thresholding feature in MIM Software. Each physician was asked to adjust the contour based on the 120 kVp, 50 keV, and MRI image set, sequentially, to mimic the utilization of scans in a clinical workflow. Intra- and inter-physician variability were compared using several metrics (Dice coefficient, mean distance to agreement, Hausdorff distance, and contour volume). Inter-physician variability was measured by comparing contours created on the same image set by different physicians. Intra-physician variability was measured by comparing contours created with different image sets for the same physician. Qualitative ratings were assigned for each of the 16 patients based on tumor conspicuity, edge sharpness, contouring confidence, and overall image quality.

Results

Optimization of monoenergetic energy

Contrast-to-noise assessment

As demonstrated in Figure 1a, b, an increase in image contrast between tumor GTV and healthy tissue was measured as the VMI energy decreased. The GTV contrast increased from 4.44 ± 10.0 HU for the 120 kVp image to 25.6 ± 22.3 HU (P = 1 × 10^-6) and 42.0 ± 32.9 HU (P = 1 × 10^-6) for the 50 keV and 40 keV images, respectively. This corresponds to an 846% (P = 1 × 10^-6) and 477% (P = 1 × 10^-6) increase in contrast for the 40 keV and 50 keV images, respectively. The increase in contrast is accompanied by an increase in noise. The 50 keV and 40 keV VMIs also demonstrated an increase in noise by 46% (P = 4 × 10^-11) and 93% (P = 4 × 10^-12), respectively, relative to the 120 kVp-equivalent image (see Fig. 1c, d).

Figure 1.

(a, b) Boxplots of gross tumor volume (GTV) contrast. (c, d) image noise, and (e, f) contrast-to-noise ratio (CNR). Figures on the left (a, c, and e) display values from 120 kVp-equivalent images (120 kVp), virtual monoenergetic images (VMI) at 50 keV (50 keV), and VMI at 40 keV (40 keV). Figures on the right (b, d, and e) display the increase in (b) GTV contrast, (d) image noise, and (f) CNR for each patient with respect to the 120 kVp image, maintaining the patient relationship.

Overall, the gain in contrast at lower energies outweighed the increase in noise, and the average CNR values increased from 0.54 ± 1.16 for the 120 kVp image to 1.99 ± 1.64 (P = 3 × 10^-7) for the 50 keV image and 2.51 ± 1.88 (P = 1 × 10^-6) for the 40 keV image. The increase in CNR for low-energy VMIs is plotted in Figure 1e, f, and visually demonstrated in Figure 2a-c.

Figure 2.

Images showing regions of interest (ROIs) placed in the gross tumor volume (red contour) and in the geniohyoid muscle (blue contour) for the (a) 120 kVp-equivalent image, (b) 50 keV monoenergetic image, (c) 40 keV monoenergetic image, and (d) average ranking of the virtual monoenergetic image energy by the physicians, with 1 being most favorable and 5 being least.

Physician preference

Optimal VMI energy was also assessed based on physician preference. Figure 2d displays the average ranking for each VMI energy by the physicians, with 1 being the most favorable ranking and 5 being the least. The 45 keV and 50 keV images were ranked the lowest (most favorable) with values of 2.0 + 1.3 and 2.1 + 0.9, respectively. The variation in rank across images was larger for the 45 keV image than for the 50 keV image, highlighting that while the high CNR at low-energy VMIs may be ideal for contouring, increases in noise, motion, and metal artifacts found in low-energy VMIs can be a hindrance. For consistency in image quality, the 50 keV image was selected for further contouring studies.

Contouring study

Qualitative measurements

Each image set was rated for perceived tumor conspicuity, tumor edge sharpness, contouring confidence, and overall image quality with the addition of each new image set. The qualitative results, shown in Figure 3a, demonstrate an increase in ratings for each category with the addition of the 50 keV VMI and the MRI image across all physicians. For each category, the largest increase was found with the addition of 50 keV VMI. The largest increase was found for the perceived tumor edge sharpness with the addition of the 50 keV image (increase of 1.86 ± 0.27).

Figure 3.

(a) Histogram plot demonstrating an increase in perceived tumor conspicuity, tumor edge sharpness, contouring confidence, and overall image quality with the addition of each new image set. Measurements corresponding to the addition of the 120 kVp, 50 keV virtual monoenergetic image (VMI), and magnetic resonance images (MRI) are shown in blue, orange, and gray, respectively. Example of a (b) 120-kVp image, (c) 50 keV VMI, and (d) MRI of a patient with base of tongue disease. The corresponding contours (blue: 120 kVp, green: 50 keV VMI, and red: MRI) are displayed on a 50 keV VMI for (e-g) physicians 1-3.

Quantitative measurements

An example of the 3 image types along with the corresponding contours from each physician can be seen in Figure 3b-g. The addition of the 50 keV VMI to the 120 kVp image resulted in a statistically significant increase in contour volume. The average contour volume continued to increase with the inclusion of the MRI (V [120 kVp] = 24.01 cm³, V [50 keV VMI] = 25.8 cm³, and V [MRI] = 29.4 cm³; P = 1.6 × 10^-6 and P = 1.4 × 10^-14). The average percentage increase in volume between the 120 kVp image and 50 keV image was 7.4%, and between the 120 kVp image and MRI was 22.3%. Figure 4 shows the increase in contour volume with the addition of each image data set.

Figure 4.

Histogram plots demonstrating (a) the volume of each contour created and (b) the increase in volume from the 120 kVp-derived contour with the addition of each image to the contouring session. The contours created with the addition of the 120 kVp, 50 keV virtual monoenergetic images, and the magnetic resonance images are shown in blue, orange, and gray, respectively.

Assessment of intra-physician variability demonstrated the greatest agreement between contours created with both CT image sets, the 120 kVp and 50 keV VMI. However, better agreement was measured between the contours created with the MRI image set and the 50 keV VMI as compared with the contours created with the MRI image set and 120 kVp image set. Figure 5a-d displays the Hausdorff distance, mean distance to agreement, Dice coefficient, and Jaccard coefficient for each contour averaged across all patients and displayed for each physician.

Figure 5.

(a-d) Histogram plots demonstrating similarity metrics comparing intraphysician variability for each contour. Data demonstrating the agreement in contours with the 120 kVp image and 50 keV virtual monoenergetic image (VMI) are shown in blue, the 50 keV VMI and the magnetic resonance image (MRI) in orange, and the 120 kVp and MRI images in gray. (e-h) Histogram plots demonstrating similarity metrics comparing interphysician variability for each contour within a given image set. Data for the 120 kVp, 50 keV VMIs, and the MRIs are shown in blue, orange, and gray, respectively.

Inter-physician variability showed no statistical change between the addition of the 50 keV VMI and 120 kVp image alone. Inter-physician variability increased with the addition of MRI. As shown in Figure 5e-h, the largest deviations were seen with average Hausdorff distance (HD [120 kVp] = 8.62, HD [50 keV VMI] = 9.01, and HD [MRI] = 10.63; P = .2 and P = .04).

Discussion

The results of this study demonstrate that the inclusion of 50 keV VMIs alongside standard 120 kVp CT images in treatment planning leads to significant alterations in GTV delineation. There was a clear improvement in CNR demonstrated with DECT that likely contributed to these changes. Notably, the addition of the 50 keV VMI resulted in an increase in GTV volume. This volumetric expansion contributed to improved agreement between CT-derived contours and those based on MRI, as evidenced by enhanced similarity metrics when comparing 50 keV VMI-derived GTVs to MRI-derived GTVs, relative to those derived from 120 kVp CT alone. These findings suggest that the inclusion of the 50 keV image improved perceived tumor conspicuity, edge sharpness, and contour confidence, and that the regions of contour expansion are often corroborated by corresponding findings on MRI. This may have implications for patients treated in centers without ready access to MRI but with a modern radiation therapy planning CT capable of DECT, though further study is needed.

Although the use of the 50 keV VMI and the MRI substantially impacted the GTV contours, the introduction of these additional image sets decreased inter-observer agreement. The increase in volume likely reflects the greater visibility of contrast uptake due to increased signal-to-noise ratio with MRI and 50 keV VMI compared with the 120 kVp CT images. The increased variability with MRI is likely due to a variety of factors, including variability of training, experience with MRI, and lack of clear guidelines on use of MRI in contouring. It is likely that some participants interpreted the increased enhancing area seen on MRI as tumor, whereas some interpreted it as due in part to inflammation, given prior confidence in 120 keV-based contours. An elevated inter-observer variability was also seen in the R-IDEAL consortium’s work and was attributed to lack of contouring guidelines for MRI.³ Thus, additional guidelines are needed for uniform incorporation of MRI and low-energy VMI CT images.

It is also noteworthy that the impact of the 50 keV VMI on GTV delineation varied by anatomic subsite. Figure 6 presents the average percentage increase in GTV volume with the addition of the 50 keV VMI and MRI, stratified by subsite: base of tongue (BOT) only, BOT and tonsil, tonsil only, and nasopharynx lesions. The volumetric increase with the addition of the 50 keV image set was most prominent in tumors involving the BOT and tonsil regions, with an average increase of 4 cm³ or 11.2%, compared with an average increase of 0.8 cm³ or 3.5% in nasopharyngeal tumors (data shown in Table E1). The relatively smaller change for the nasopharynx sites is likely attributable to the proximity of these tumors to dense bony structures, which reduce the incremental benefit of low-energy VMI-derived contrast enhancement. In clinical practice, we find the largest impact of MRI and low-energy VMI images in the BOT, where anterior extension of tumor can sometimes only be seen on these modalities.

Figure 6.

Histogram plots showing the average percentage increase in gross tumor volume with the addition of the 50 keV virtual monoenergetic image and magnetic resonance image to the 120 kVp planning CT image, stratified by disease subsite: base of tongue (BOT) only, BOT and tonsil, tonsil only, and nasopharynx.

Similarity metrics comparing GTVs derived from the 50 keV VMI and MRI showed consistently better agreement than those comparing 120 kVp CT and MRI. This trend held across all anatomic subsites, with the exception of the Hausdorff distance in nasopharyngeal cases. For these cases, the average Hausdorff distance between the 50 keV VMI and MRI was not significantly different from that between the 120 kVp CT and MRI (HD [M50-MRI] = 7.42 and HD [C120-MRI] = 7.32; P = .41). These findings further suggest that the utility of low-energy VMIs may be reduced in anatomic regions surrounded by osseous structures, such as skull base tumors. This agrees with the previous work of Ng et al,¹² showing the lowest interobserver agreement for skull base tumors with MRI and for nonskull base tumors with DECT. Additional postprocessed images derived from DECT may provide more benefit for nasopharynx tumors. Effective atomic number, noncalcium, and iodine maps can be exploited to visualize and characterize bone marrow^18,19 and, as a result, have been shown to be more accurate in the diagnosis of skull base invasion when compared with SECT and MRI for nasopharynx tumors.²⁰ Although the scope of this study is limited to the effect of low-energy VMIs on tumor segmentation, a larger study investigating the effects of additional DECT reconstructions is needed to assess the complete benefits of DECT in HN tumor delineation.

Compared with the only previously published study assessing inter-observer variability for DECT VMIs (Ng et al¹²), the inter-observer agreement in this investigation was higher. This discrepancy likely stems from differences in study design. Although Ng et al evaluated variability using low-energy VMIs in isolation, the present study assessed the impact of adding VMIs to an imaging set with an established GTV contour, mimicking clinical practice and providing a robust assessment of the additive value of DECT. The decrease in variability with this study design provides an important insight by suggesting that each study provides unique information, leading to changes in contours. It also suggests that DECT does not completely substitute for MRI in the HN workflow, which is consistent with our clinical experience. Further study is needed to assess the functional consequences of these changes in terms of tumor control and toxicity.

Although the focus of this work is DECT-derived VMIs, several publications have recently presented deep learning-generated VMIs from SECTs.21, 22, 23 These DL-generated VMIs have shown improvement in image quality compared with SECT images²³; however, more work is needed to demonstrate their applicability.

There were several limitations to this study. Each diagnostic PET and MRI was rigidly registered to the CT images. Each image registration induces uncertainty into the contouring process. Although patient selection was limited to patients with similar setups between image data sets, immobilization was not used for diagnostic PET and MRI; therefore, minor differences in setup, such as neck flexion, jaw position, and head tilt, were inherent in image registrations. Due to the sequential workflow of this study, carryover or anchoring bias could influence the results. This is also a limited scope study including only 16 patients. Another limitation is that all authors are from the same institution, where we share a contouring philosophy and training. A large, multi-institutional study is needed to further characterize the utility of DECT.

Conclusions

This study demonstrates that incorporating 50 keV VMIs into standard HN radiation therapy planning may lead to larger GTVs that are more consistent with findings on MRI. Quantitative and qualitative analyses confirmed that low-energy VMIs enhance perceived tumor conspicuity and edge sharpness, leading to increased contour volumes and improved agreement with MRI-based contours. While the most substantial impact was observed at disease sites surrounded by soft tissues, such as the BOT and tonsil, the benefit was less pronounced in regions adjacent to bony anatomy, such as the nasopharynx. Overall, this work supports the integration of low-energy VMIs into multimodal imaging workflows and highlights their value in refining radiation therapy planning for HN cancers. It also suggests a need for a large, multi-institutional study and the development of guidelines for more uniform incorporation of low-energy VMI and MRI into treatment volumes.

Disclosures

Patrick Wohlfahrt and Jainil Shah are employees of Siemens Healthineers. The authors also disclose the following activities that are unrelated to this work. Brett Morris received funding from Icotec for conference support and serves as Vice Chair of the Epic Radiation Oncology Steering Board. Michael Lawless and Adam Burr received research funding from GE Healthcare. Grace Blitzer received payment for lectures at Meriter Hospital for grand rounds on radiation therapy for osteoarthritis and has stock disclosures for Quench Cellular Therapeutics as co-founder. Jessica Schuster received payment as an ACRO HN educational speaker, received funding support for attending the Varian Functional Radiation Advisory Meeting and ACRO, and serves on the ACRO board of directors. Randy J. Kimple received consulting fees from Guidepoint Global and HunaTek, has patent disclosures through Wisconsin Alumni Research Foundation, and has stock disclosures for Quench Cellular Therapeutics with stock and ownership, member of the board, and CEO.