Abstract
Determining whether input features are consistently strongly correlated with organ-at-risk (OAR) dose across different diseases can improve the accuracy of dose-volume histogram (DVH) prediction in knowledge-based treatment planning. In this study, we utilized patient geometric and dosimetric data to establish correlations with OAR doses and investigated the differences between distance-to-dose and dosimetric-to-dose relationships. Ninety nasopharyngeal cancer radiotherapy plans, 66 esophageal cancer radiotherapy plans and 52 rectal cancer radiotherapy plans were selected as study objects. Geometric feature was quantified by the distance-to-target histogram (DTH), and dosimetric feature was quantified by the conformal-plan-dose-volume histogram (CPDVH). The distance-to-dose correlation and dosimetric feature-to-dose correlation were calculated. DTH-DVH and CPDVH-DVH correlations were calculated for fractional OAR volumes of 30%, 50% and 60%. The correlations were calculated using the Pearson product-moment correlation coefficient (R). Compared to the distance-to-dose correlation, dosimetric feature achieved a stronger correlation (R) with OAR doses for 10 out of 13 OARs, including those to the brainstem (− 0.83, 0.93) and lung (− 0.60, 0.92). Compared to the DTH-DVH correlation, the CPDVH dose-volume achieved a stronger correlation (R) with the DVH dose-volume for 10 out of 13 OARs, including the brainstem (mean, − 0.67 vs. 0.80) and lung (mean, − 0.64 vs. 0.76). Compared to the CPDVH-DVH correlation, the DTH distance-volume showed a stronger correlation with the DVH dose-volume at the parotid (mean, − 0.88 vs. 0.60) and bladder (mean, − 0.56 vs. 0.46). The results indicated that the dosimetric feature showed a stronger correlation with most OAR doses than the distance-to-dose correlation, which is helpful for accurate DVH prediction. Patient geometry influences dosimetric feature-to-dose correlation when many OARs voxels are located inside the C-shaped target region or when the OARs significantly overlap with the target, and using only dosimetric feature for the DVH prediction of these OARs is not appropriate.
Keywords: Dosimetric feature, Correlation analysis, Dose-volume histogram prediction, Nasopharyngeal cancer, Esophageal cancer, Rectal cancer
Subject terms: Cancer therapy, Head and neck cancer, Lung cancer
Introduction
As an effective oncotherapy method, radiotherapy is widely employed in the treatment of various diseases1–3. Common radiotherapy techniques include 3D conformal radiation therapy, intensity modulated radiation therapy (IMRT) and rotational volumetric modulated arc therapy (VMAT). Before radiation therapy, a clinically acceptable treatment plan must be created using optimal dose-volume constraints, which are typically unknown beforehand4. However, manual adjustments often result in suboptimal plans, with significant quality variations due to differences in planners’ expertise. To address these challenges, several studies have proposed patient geometry-driven organ-at-risk (OAR) dose prediction methods based on geometric and dosimetric features from prior patients5–8. Among these, geometric feature-based prediction has been widely adopted to establish reliable dose objectives, reducing variability in plan quality and enhancing overall consistency9–15.
The representative patient geometric feature descriptors are the overlap volume histogram (OVH) and the distance-to-target histogram (DTH), which quantify the spatial relationship between OARs and the target16–18. These metrics enable the establishment of correlative relationships between OVH/DTH and the dose-volume histogram (DVH) for DVH prediction. The accuracy of DVH prediction relies heavily on the strength of the correlation between OVH/DTH and DVH19. For instance, one study demonstrated that incorporating the in-field OAR volume—which exhibits a strong correlation with OAR dose for both the bladder and rectum—can further improve OVH-based DVH prediction19. In recent years, most commercially available knowledge-based planning (KBP) systems have integrated patient geometry-driven DVH prediction to optimize treatment plan design11–14,20. However, it remains unclear whether geometric features alone can fully capture dose variations, and few studies have systematically explored this limitation.
Previous studies have demonstrated that multiple factors may influence OAR DVHs19,21,22. For instance, the resultant DVH distribution can depend on both the spatial relationship between OARs and the radiation field boundaries21 and the irradiated OAR volume within the treatment field22. These studies reflect the difficulty of explaining the existence of DVH differences among patients with similar geometric characteristics using DVH prediction models based on patient geometric features. Features strongly correlated with OAR DVH can aid in DVH prediction. Consequently, these studies have explored various factors influencing OAR dose and validated their predictive role in DVH modeling19,21,22. DVH prediction is also influenced by the choice of prediction model. Mainstream approaches include density estimation-based DVH prediction3, regression modeling of linear distance-dose relationships23, and neural network-based methods24.
Over the past five years, numerous studies have successfully incorporated dosimetric characteristics into OAR dose prediction models, demonstrating promising results25–29. However, it remains unclear whether a strong correlation exists between dosimetric features and OAR doses across all OARs in representative cases. In this study, we identified dosimetric features most relevant to dose prediction and conducted correlation analyses between (1) patient geometric features and OAR dose, and (2) patient dosimetric features and OAR dose. Our findings are expected to enhance the understanding of differences between distance-to-dose and dosimetric feature-to-dose correlations, thereby facilitating more appropriate application of both geometric and dosimetric features in DVH prediction.
We first introduce the method of selecting prior plans, which were used as the database. Considering the differences in geometric feature for different diseases, three representative diseases, i.e., nasopharyngeal cancer of the head and neck, esophageal cancer of the chest, and rectal cancer of the abdomen, were selected as the study subjects. Subsequently, we detail the patient geometric and dosimetric features used in this study. Finally, we present and discuss the resulting correlations between (1) geometric features and OAR dose, and (2) dosimetric features and OAR dose.
Materials and methods
Patient dataset
A total of 90 nasopharyngeal cancer clinical radiotherapy plans, 66 esophageal cancer clinical radiotherapy plans and 52 rectal cancer clinical radiotherapy plans were selected as the database. The nasopharyngeal cancer clinical radiotherapy plans were generated using 9-field 8-MV photon beams. The beam gantry angles were: 160°, 120°, 80°, 40°, 0°, 200°, 240°, 280° and 320°. The esophageal cancer clinical radiotherapy plans were generated using 5-field 8-MV photon beams. The beam gantry angles were: 160°, 40°, 0°, 200° and 320°. The rectal cancer clinical radiotherapy plans were generated using 5-field 8-MV photon beams with gantry angles of 150°, 75°, 0°, 220° and 290°. The radiotherapy plans were generated using the MONACO TPS (Elekta CMS, Maryland Heights, MO, version 6.0).
The target prescription dose of the nasopharyngeal cancer plan was 70 Gy in 32 fractions. Three planning target volumes (PTVs), i.e., PTV70, PTV60 and PTV54, were specified in the nasopharyngeal cancer IMRT plan. For esophageal cancer, the target prescription dose was 60 Gy in 30 fractions. Three PTVs (PTV60, PTV54 and PTV50) were specified in the esophageal cancer IMRT plan. For rectal cancer, the target prescription dose was 50 Gy in 25 fractions. Two PTVs (PTV50 and PTV45) were specified in the rectal cancer IMRT plan. The detailed dose optimization constraints required when generating a radiation therapy plan in our center are listed in Table 1.
Table 1.
The dose optimization constraints required when generating a radiation therapy plan.
| Nasopharyngeal cancer | Esophageal cancer | Rectal cancer | |||
|---|---|---|---|---|---|
| Structure | Constraint | Structure | Constrain | Structure | Constraint |
| PTV70 | D0 ≤ 75 Gy | PTV60 | D0 ≤ 65 Gy | PTV50 | D0 ≤ 55 Gy |
| - | D100 ≥ 71 Gy | - | D95 ≥ 60.5 Gy | - | D97 ≥ 51 Gy |
| - | D97 ≥ 72 Gy | - | - | - | - |
| PTV60 | D100 ≥ 60 Gy | PTV54 | D95 ≥ 54.5 Gy | PTV45 | D97 ≥ 46 Gy |
| PTV54 | D100 ≥ 54 Gy | PTV50 | D95 ≥ 50.5 Gy | - | - |
| Brainstem | D0 ≤ 54 Gy | Bilateral lungs | V20 ≤ 37% | Bladder | V50 ≤ 50% |
| Spinal cord | D0 ≤ 45 Gy | - | - | - | V60 ≤ 20% |
| Bilateral optic nerves | D0 ≤ 54 Gy | - | - | Colon | D0 ≤ 52 Gy |
| Bilateral lens | D0 ≤ 9 Gy | - | - | - | - |
| Bilateral parotids | V30 ≤ 50% | - | - | Small intestine | D0 ≤ 55 Gy |
| Bilateral eyes | D0 ≤ 50 Gy | Heart | V60 ≤ 33% | - | V50 ≤ 5% |
| Bilateral temporal lobes | D0 ≤ 72 Gy | - | V45 ≤ 67% | Bilateral femoral heads | V50 ≤ 5% |
Vx (volume receiving greater than x Gy), Dy (the dose at percentage volume y).
Geometric feature and dosimetric feature
The patient’s geometric features were quantified using the DTH, a well-established descriptor for evaluating spatial relationships between OARs and PTVs. The detailed mathematical formulation of DTH was previously described by Zhu et al.18. Distance calculations from each voxel to the reference PTV were performed using our in-house MATLAB algorithm (version R2021a). For this study, we defined the reference PTVs as follows: PTV54 for nasopharyngeal cancer cases, PTV50 for esophageal cancer cases, and PTV45 for rectal cancer cases.
The patient dosimetric feature was derived from a manually generated conformal plan. The OAR dosimetric feature was quantified using the conformal-plan-dose-volume histogram (CPDVH). The process of generating a conformal plan is as follows:
For nasopharyngeal cancer patients, conformal plans were generated for 9-field 8-MV photon beams with gantry angles of 160°, 120°, 80°, 40°, 0°, 200°, 240°, 280°, and 320°. Each field was fitted to PTV54. The slice thickness of the CTs was 0.3 cm. For esophageal cancer patients, the conformal plans were generated for 5-field 8-MV photon beams with gantry angles of 160°, 40°, 0°, 200° and 320°. Each field was fitted to PTV50. The slice thickness of the CTs was 0.5 cm. For rectal cancer patients, conformal plans were generated for 5-field 8-MV photon beams with gantry angles of 150°, 75°, 0°, 220° and 290°. Each field was fitted to the PTV45. The slice thickness of the CTs was 0.3 cm. The conformal plan was generated using MONACO TPS.
The correlation analysis
After finishing the extraction process for the geometric and dosimetric feature, each OAR voxel was characterized by three parameters: (1) the distance to the PTV surface (r); (2) the conformal plan dose (D*); and (3) the IMRT dose (D). The relationships between the distance and the IMRT dose and between the conformal plan dose and the IMRT dose could be established. For detailed analysis, we focused on three clinically relevant OARs: the eye (nasopharyngeal cancer cases), lung (esophageal cancer cases), and small intestine (rectal cancer cases).
Figure 1 presents representative examples of DTH, CPDVH and DVH. A DTH-DVH relationship was established (as illustrated in Fig. 1.a). For the same fractional volume v, a distance r corresponds to a dose D. Figure 1.b illustrates the CPDVH-DVH relationship. For the same fractional volume v, a dose D* corresponds to a dose D. A correlation analysis was calculated using the Pearson product-moment correlation coefficient (R), which was also used for the correlation analysis in reference19. The Pearson correlation coefficient quantifies the strength and direction of linear association between two variables, with possible values ranging from − 1 to 1. A coefficient of 1 represents a perfect positive linear relationship, −1 indicates a perfect negative linear relationship, and 0 denotes the absence of linear correlation. For this analysis, we computed correlation coefficients using MATLAB’s built-in corrcoef function (version R2021a, MathWorks, Natick, MA).
Fig. 1.
An example of the DTH-DVH relationship (a) and the CPDVH-DVH relationship (b) for a certain OAR. For the same fractional volume v, a distance r corresponds to a dose D (rv-to-Dv). For the same fractional volume v, a conformal plan dose D* corresponds to a dose D (D*v-to-Dv).
Results
The distance-to-dose correlation and the dosimetric feature-to-dose correlation
Figure 2 demonstrates the relationships between (1) distance-to-PTV and IMRT dose, and (2) conformal plan dose and IMRT dose, with each scatter point representing individual voxels within the analyzed OAR. The analysis reveals two distinct patterns: First, an inverse relationship exists between distance-to-PTV and IMRT dose, where dose levels decrease rapidly with increasing distance from the PTV surface. Second, a strong positive correlation is observed between conformal plan dose and IMRT dose, indicating that higher conformal plan doses correspond to proportionally higher IMRT doses.
Fig. 2.
The correlative relationship between the distance to the PTV surface and dose of IMRT plan and between the dose of conformal plan and the dose of IMRT plan. The scatter point represents the voxel inside a certain OAR.
The distance-to-dose correlation and dosimetric feature-to-dose correlation results of each OAR are illustrated in Table 2. In nasopharyngeal cancer, compared with the distance-to-dose correlation (R), the dosimetric feature correlation with the IMRT dose was stronger for five of the seven OARs, including the brainstem (− 0.83 vs. 0.93) and lens (− 0.82 vs. 0.95). The distance to the PTV surface correlation with the IMRT dose was stronger than the dosimetric feature-to-dose correlation at the spinal cord (− 0.72 vs. 0.66) and parotid (− 0.83 vs. 0.70). For esophageal cancer, compared with the distance-to-dose correlation (R), the dosimetric feature correlation with the IMRT dose was stronger for all OARs, including the lung (− 0.60 vs. 0.92) and heart (− 0.61 vs. 0.85). For rectal cancer, compared with the distance-to-dose correlation (R), the dosimetric feature correlation with the IMRT dose was stronger for three of the four OARs, including the colon (− 0.83 vs. 0.95), small intestine (− 0.69 vs. 0.94) and femoral head (− 0.86 vs. 0.93). The distance to the PTV surface correlation with the IMRT dose was stronger than the dosimetric feature-to-dose correlation at the bladder (− 0.90 vs. 0.76).
Table 2.
The average pearson correlation coefficients between the distance to the PTV surface and IMRT plan dose and between the conformal plan dose and IMRT plan dose.
| Nasopharyngeal cancer | Esophageal cancer | Rectal cancer | ||||||
|---|---|---|---|---|---|---|---|---|
| Structure | Dis-D | D*-D | Structure | Dis-D | D*-D | Structure | Dis-D | D*-D |
| Brainstem | −0.83 | 0.93 | Bilateral lung | −0.6 | 0.92 | Bladder | −0.9 | 0.76 |
| Spinal cord | −0.72 | 0.66 | ||||||
| Bilateral optic nerve | −0.9 | 0.94 | Colon | −0.83 | 0.95 | |||
| Bilateral lens | −0.82 | 0.95 | ||||||
| Bilateral parotid | −0.83 | 0.7 | Small intestine | −0.69 | 0.94 | |||
| Bilateral eye | −0.74 | 0.96 | ||||||
| Bilateral temporal lobe | −0.84 | 0.97 | Heart | −0.61 | 0.85 | Bilateral femoral head | −0.86 | 0.93 |
Dis represents the distance to the PTV surface; D represents the IMRT plan dose; D* represents the conformal plan dose.
The DTH-DVH correlation and the CPDVH-DVH correlation
Figure 3 shows the DTH-DVH correlations and the CPDVH-DVH correlations for dose volumes of certain OARs of 30%, 50% and 60%. The Pearson correlation coefficients between DTH distance volumes and DVH dose volumes and between CPDVH dose volumes and DVH dose volumes are listed in Table 3. For nasopharyngeal cancer, compared with the DTH-DVH correlation (R), the CPDVH dose-volume correlation with the DVH dose-volume was stronger for five of the seven OARs, including the brainstem (mean, − 0.67 vs. 0.80) and lens (mean, − 0.62 vs. 0.85). The DTH distance-volume correlation with the DVH dose-volume was stronger than the CPDVH-DVH correlation at the parotid (mean, − 0.88 vs. 0.60).
Fig. 3.
The DTH-DVH correlations for 30%, 50% and 60% dose volumes of the eye, lung and small intestine; the CPDVH-DVH correlations for 30%, 50% and 60% dose volumes of the eye, lung and small intestine. rx: distance at x% volume in the DTH curve. D*x: conformal plan dose at x% volume in the CPDVH curve. Dx: dose at x% volume in the DVH curve. For nasopharyngeal cancer, the DTH-DVH correlations and the CPDVH-DVH correlations were established based on 90 patients. For esophageal cancer, the DTH-DVH correlations and the CPDVH-DVH correlations were established based on 66 patients. For rectal cancer, the DTH-DVH correlations and the CPDVH-DVH correlations were established based on 52 patients.
Table 3.
Pearson correlation coefficients between the DTH distance volumes and DVH dose volumes and pearson correlation coefficients between the CPDVH dose volumes and DVH dose volumes.
| Nasopharyngeal cancer | Esophagus cancer | Rectal cancer | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Structures | DTH | CPDVH | Structures | DTH | CPDVH | Structures | DTH | CPDVH | Structures | DTH | CPDVH |
| Brainstem | Bilateral parotids | Bilateral lungs | Bladder | ||||||||
| D30 | −0.71 | 0.75 | D30 | −0.92 | 0.7 | D30 | −0.67 | 0.7 | D30 | −0.65 | 0.4 |
| D50 | −0.65 | 0.82 | D50 | −0.86 | 0.54 | D50 | −0.64 | 0.84 | D50 | −0.54 | 0.48 |
| D60 | −0.64 | 0.84 | D60 | −0.85 | 0.56 | D60 | −0.6 | 0.74 | D60 | −0.48 | 0.49 |
| Mean | −0.67 | 0.8 | Mean | −0.88 | 0.6 | Mean | −0.64 | 0.76 | Mean | −0.56 | 0.46 |
| Spinal cord | Bilateral eyes | - | Colon | ||||||||
| D30 | −0.2 | 0.3 | D30 | −0.64 | 0.84 | - | - | - | D30 | −0.86 | 0.92 |
| D50 | −0.24 | 0.15 | D50 | −0.68 | 0.89 | - | - | - | D50 | −0.74 | 0.91 |
| D60 | −0.56 | 0.53 | D60 | −0.68 | 0.88 | - | - | - | D60 | −0.67 | 0.9 |
| Mean | −0.33 | 0.33 | Mean | −0.67 | 0.87 | - | - | - | Mean | −0.76 | 0.91 |
| Bilateral optic nerves | Bilateral temporal lobes | Heart | Small intestine | ||||||||
| D30 | −0.85 | 0.96 | D30 | −0.65 | 0.94 | D30 | −0.85 | 0.9 | D30 | −0.79 | 0.96 |
| D50 | −0.83 | 0.94 | D50 | −0.48 | 0.91 | D50 | −0.85 | 0.96 | D50 | −0.7 | 0.92 |
| D60 | −0.82 | 0.93 | D60 | −0.44 | 0.88 | D60 | −0.84 | 0.95 | D60 | −0.68 | 0.9 |
| Mean | −0.83 | 0.94 | Mean | −0.52 | 0.91 | Mean | −0.85 | 0.94 | Mean | −0.72 | 0.93 |
| Bilateral lens | - | - | Bilateral femoral heads | ||||||||
| D30 | −0.42 | 0.87 | - | - | - | - | - | - | D30 | −0.5 | 0.64 |
| D50 | −0.72 | 0.81 | - | - | - | - | - | - | D50 | −0.51 | 0.66 |
| D60 | −0.72 | 0.86 | - | - | - | - | - | - | D60 | −0.48 | 0.68 |
| Mean | −0.62 | 0.85 | - | - | - | - | - | - | Mean | −0.5 | 0.66 |
Dy (the dose to the highest y% of volume.
Regarding esophageal cancer, the CPDVH dose-volume correlation with the DVH dose-volume was stronger than the DTH-DVH correlation in the lung (mean, − 0.64 vs. 0.76) and heart (mean, − 0.85 vs. 0.94).
For rectal cancer, the CPDVH dose-volume correlation with the DVH dose-volume was stronger than the DTH-DVH correlation for three of the four OARs, including the colon (mean, − 0.76 vs. 0.91), small intestine (mean, − 0.72 vs. 0.93) and femoral head (mean, − 0.50 vs. 0.66). The DTH distance-volume correlation with the DVH dose-volume was stronger than the CPDVH-DVH correlation at the bladder (mean, − 0.56 vs. 0.46).
Discussion
In this study, we established and investigated a correlative relationship between patient dosimetric features and OAR doses. The results indicate that, for most OARs, the dosimetric feature exhibits a strong correlation with the OAR dose (R > 0.90).
The average dosimetric feature-to-dose correlation coefficient was below 0.80 for the spinal cord, parotids, and bladder (Table 2). This discrepancy arises because portions of these OARs—such as the spinal cord—are adjacent to the C-shaped target volume, while others (e.g., the parotid and bladder) partially overlap with the target (Fig. 4). For OAR voxels overlapping with the PTV or located within the C-shaped target region, the dose fall-off rate in the unoptimized conformal plan is relatively low due to its fixed beam intensity. In contrast, the clinical IMRT plan achieves a sharper dose gradient at the PTV boundary to spare OARs (Fig. 4). Consequently, even minor dose variations in the conformal plan may correspond to significant dose changes in the IMRT plan, thereby weakening the correlation between dosimetric features and OAR doses.
Fig. 4.
The dose distributions of the IMRT plan (a) and conformal plan (b) in a nasopharyngeal cancer patient. The dose distributions of the IMRT plan (c) and conformal plan (d) in a rectal cancer patient. The blue arrow represents the direction of the dose fall-off. Abbreviations: SC (spinal cord).
The dosimetric feature used in this study differs from that employed by Ma et al.26. In their work, the dosimetric feature—derived from a PTV-only optimized plan—represented the target dose distribution without incorporating OAR dose constraints. Generating such a plan requires a dedicated dose optimization process26. In contrast, our study utilized an unoptimized dosimetric feature obtained through a simpler methodology. Future research could explore the correlation between the PTV-only plan’s dosimetric feature and OAR doses, comparing its feature-to-dose relationship with that of the conformal plan to assess potential differences.
Wall et al. reported that the use of factors that are strongly correlated with the OAR dose helps to accurately predict the DVH19. This study helps us to understand the relationships between patient geometric feature and OAR dose and between patient dosimetric feature and OAR dose, which can lead to the appropriate use of patient geometric and dosimetric feature when predicting DVH. The use of dosimetric feature for predicting DVH is recommended when the dosimetric feature-to-dose correlation is stronger than the distance-to-dose correlation, while the use of geometric feature for predicting DVH is appropriate when the distance-to-dose correlation is stronger than the dosimetric feature-to-dose correlation.
To assess the influence of input features on DVH prediction accuracy, we developed two distinct prediction models: (1) a patient geometric feature-based model (DTH method) and (2) a patient dosimetric feature-based model (CPDVH method). Figure 5 compares the DVH prediction errors for various OARs between these two approaches. The results demonstrate that the dosimetric feature-based approach yielded superior prediction accuracy for the lung and small intestine, whereas the geometric feature-based method performed better for the spinal cord. This discrepancy may be attributed to the relatively weak dosimetric feature-to-dose correlation observed for the spinal cord (R = −0.72 vs. 0.66). Furthermore, the CPDVH method showed limited capability in differentiating high-dose regions across patients, as these regions exhibited considerable similarity in the CPDVHs of different patients. This limitation likely contributes to the method’s reduced predictive performance for certain OARs.
Fig. 5.
DVH prediction error (for the DTH method, δV = VDTH_pred(D)-VTPS(D); for the CPDVH method, δV = VCPDVH_pred(D)-VTPS(D)) of the selected OARs in 20 nasopharyngeal cancer test cases, 15 esophageal cancer test cases and 15 rectal cancer test cases.
The patient dosimetric feature-based DVH prediction model cannot accurately predict the DVHs of OARs; to address this issue, the dosimetric feature of different fields can possibly be separated. Even in the complex C-shaped target region, the dose fall-off information of different fields, rather than the uniform dose distribution from the dose summation of different fields, should exist.
In this study, a small dataset of 52 rectal cancer patients was used to study the difference between the distance-to-dose correlation and the dosimetric feature-to-dose correlation. This is because the disease cases with the same target prescription dose were very limited in our treatment center. Although our results indicated that an obvious difference between the distance-to-dose correlation and the dosimetric feature-to-dose correlation existed, a small dataset may not be sufficient to fully evaluate the difference. Future work will be focused on using more cases when more rectal cases are available.
To evaluate the feasibility and generalizability of establishing dosimetric feature-to-OAR dose correlations across different anatomical geometries, we selected three distinct disease sites—nasopharyngeal cancer, esophageal cancer, and rectal cancer—as study models. Our analysis revealed that disease-specific geometric characteristics, particularly the spatial relationships between OARs and PTVs, significantly influence the strength of correlation between conformal plan dosimetric features and OAR doses. For each disease, we conducted comprehensive distance-to-dose and dosimetric feature-to-dose correlation analyses. These comparative assessments demonstrate that anatomical configuration plays a crucial role in determining the predictive value of different dose estimation approaches. While this study provides foundational insights into three representative disease models, future research should expand this investigation to additional disease sites to further validate and generalize these findings.
Conclusion
This study evaluated distance-to-dose and dosimetric feature-to-dose correlations across three cancer types: nasopharyngeal cancer, esophageal cancer, and rectal cancer. Our analysis revealed that dosimetric features generally demonstrated stronger correlations with OAR doses than distance-based metrics for most organs at risk. However, notable exceptions were observed for the spinal cord and parotids in nasopharyngeal cancer, and the bladder in rectal cancer, where geometric relationships showed superior predictive value. These findings suggest that relying solely on dosimetric features for DVH prediction may be inadequate for certain OARs, particularly when anatomical geometry plays a dominant role in dose distribution. Patient geometry influences dosimetric feature-to-dose correlation when many OAR voxels are in the C-shaped target region or when the OARs highly overlap with the target.
Acknowledgements
This work was supported by the Natural Science Foundation of Shandong Province (Youth Program, ZR2022QA081), the Medical Research Project of Artificial Intelligence Plus of Daping Hospital (Key project # ZXAIZD004) and the Clinical Medical Research Project of Army Medical University (project # 2023XLC12).
Author contributions
Peng Zhou and Jia Luo: Methodology, Software, Supervision, Statistical analyses, Validation, Writing-original draft. Honghao Xu: Data curation and Formal Analysis. Shengxiu Jiao and Chuan Chen: Conceptualization, Funding acquisition, Project administration, Writing-review and editing and Investigation.
Data availability
The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request.
Declarations
Competing interests
The authors declare no competing interests.
Ethical statement
This work was approved by our internal ethics committee (NO: 2025(31)).
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
These authors contributed equally to this work: Peng Zhou and Jia Luo.
Contributor Information
Shengxiu Jiao, Email: jiaoshx@mail2.sysu.edu.cn.
Chuan Chen, Email: sinkriver@126.com.
References
- 1.De Rose, F. et al. Dose constraints in breast cancer radiotherapy. A critical review. Radiother Oncol.202, 110591 (2025). [DOI] [PubMed] [Google Scholar]
- 2.Ma, D. et al. Exploring the advantages of intensity-modulated proton therapy: experimental validation of biological effects using two different beam intensity-modulation patterns. Sci. Rep.10 (1), 3199 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Duan, W. et al. Dosimetric comparison of gamma knife and linear accelerator (VMAT and IMRT) plans of SBRT of lung tumours. Sci. Rep.14 (1), 22949 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Wang, J., Xi, P., Cao, R. F. & Qin, H. L. Can. A multiphase direct aperture optimization for inverse planning in radiotherapy. Nucl. Sci. Tech.26 (1), 010502 (2015). [Google Scholar]
- 5.Munter, J. S. & Sjölund, J. Dose-volume histogram prediction using density Estimation. Phys. Med. Biol.60 (17), 6923–6936 (2015). [DOI] [PubMed] [Google Scholar]
- 6.Nguyen, D. et al. A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patients from patient anatomy using deep learning. Sci. Rep.9 (1), 1076 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Qi, M. et al. Region-specific three-dimensional dose distribution prediction: a feasibility study on prostate VMAT cases. J. Radiation Res. Appl. Sci.13 (1), 485–495 (2020). [Google Scholar]
- 8.Tao, C. et al. A novel knowledge-based prediction model for estimating an initial equivalent uniform dose in semi-auto-planning for cervical cancer. Radiat. Oncol.17 (1), 151 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Zhang, Y. et al. DoseDiff: Distance-Aware diffusion model for dose prediction in radiotherapy. IEEE Trans. Med. Imaging. 43 (10), 3621–3633 (2024). [DOI] [PubMed] [Google Scholar]
- 10.Moore, K. L., Brame, R. S., Low, D. A. & Mutic, S. Experience-based quality control of clinical intensity-modulated radiotherapy planning. Int. J. Radiat. Oncol. Biol. Phys.81, 545–551 (2011). [DOI] [PubMed] [Google Scholar]
- 11.Zhang, B., Babier, A. & Ruschin, M. Chan. Knowledge-based planning for gamma knife. Med. Phys.51 (5), 3207–3219 (2024). [DOI] [PubMed] [Google Scholar]
- 12.Fjellanger, K. et al. Pettersen and L. B. Hysing. Improving knowledge-based treatment planning for lung cancer radiotherapy with automatic multi-criteria optimized training plans. Acta Oncol.62 (10), 1194–1200 (2023). [DOI] [PubMed] [Google Scholar]
- 13.Johnson, C. L. et al. Advancing knowledge-based intensity modulated proton planning for adaptive treatment of high-risk prostate cancer. Med. Dosim. 49 (1), 19–24 (2024). [DOI] [PubMed] [Google Scholar]
- 14.Liu, H., Sintay, B. & Wiant, D. A two-step treatment planning strategy incorporating knowledge-based planning for head-and-neck radiotherapy. J. Appl. Clin. Med. Phys.24 (6), e13939 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Chang, A. T. Y. et al. Comparison of planning quality and efficiency between conventional and knowledge based algorithms in nasopharyngeal cancer patients using intensity-modulated radiotherapy. Int. J. Radiat. Oncol. Biol. Phys.95, 981–990 (2016). [DOI] [PubMed] [Google Scholar]
- 16.Kazhdan, M. et al. A shape relationship descriptor for radiation therapy planning. Lecture Notes Comput. Sci.5762, 100–108 (2009). [DOI] [PubMed] [Google Scholar]
- 17.Wu, B. et al. Patient geometry-driven information retrieval for IMRT treatment plan quality control. Med. Phys.36, 5497–5505 (2009). [DOI] [PubMed] [Google Scholar]
- 18.Zhu, X. et al. A planning quality evaluation tool for prostate adaptive IMRT based on machine learning. Med. Phys.38, 719–726 (2011). [DOI] [PubMed] [Google Scholar]
- 19.Wall, P. D. H., Carver, R. L. & Fontenot, J. D. An improved distance-to-dose correlation for predicting bladder and rectum dose-volumes in knowledge-based VMAT planning for prostate cancer. Phys. Med. Biol.63, 015035 (2018). [DOI] [PubMed] [Google Scholar]
- 20.Sasaki, M. et al. Comparative analysis of two dose-volume histogram prediction tools for treatment planning in volumetric-modulated Arc therapy: A multi-planner study. Med. Dosim. 49 (3), 271–275 (2024). [DOI] [PubMed] [Google Scholar]
- 21.Petit, S. F. et al. Increased organ sparing using shape-based treatment plan optimization for intensity modulated radiation therapy of pancreatic adenocarcinoma. Radiother Oncol.102 (1), 38–44 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Wu, B. et al. Improved robotic stereotactic body radiation therapy plan quality and planning efficacy for organ-confined prostate cancer utilizing overlap-volume histogram-driven planning methodology. Radiother Oncol.112 (2), 221–226 (2014). [DOI] [PubMed] [Google Scholar]
- 23.Yang, Y. et al. An overlap-volume-histogram based method for rectal dose prediction and automated treatment planning in the external beam prostate radiotherapy following hydrogel injection. Med. Phys.40 (1), 011709 (2013). [DOI] [PubMed] [Google Scholar]
- 24.Chen, X. et al. DVHnet: A deep learning-based prediction of patient-specific dose volume histograms for radiotherapy planning. Med. Phys.48 (6), 2705–2713 (2021). [DOI] [PubMed] [Google Scholar]
- 25.Jiao, S., Chen, L., Zhu, J., Wang, M. & Liu, X. Prediction of dose-volume histograms in nasopharyngeal cancer IMRT using geometric and dosimetric information. Phys. Med. Biol.64 (23), 23NT04 (2019). [DOI] [PubMed] [Google Scholar]
- 26.Ma, M., Kovalchuk, N., Buyyounouski, M. K., Xing, L. & Yang, Y. Dosimetric features-driven machine learning model for DVH prediction in VMAT treatment planning. Med. Phys.46 (2), 857–867 (2019). [DOI] [PubMed] [Google Scholar]
- 27.Cao, W., Zhuang, Y., Chen, L. & Liu, X. Application of dose-volume histogram prediction in biologically related models for nasopharyngeal carcinomas treatment planning. Radiat. Oncol.15 (1), 216 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Jiao, S., Wang, M., Chen, L. & Liu, X. Evaluation of dose-volume histogram prediction for organ-at risk and planning target volume based on machine learning. Sci. Rep.11, 3117 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Jiao, S.X., Wang, M.L., Chen, L.X. & Liu, X.W. Knowledge-based planning using both the predicted DVH of organ-at risk and planning target volume. Med. Eng. Phys.110, 103803 (2022). [DOI] [PubMed]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request.