This retrospective analysis included 12 newly diagnosed, pathologically confirmed stage I - IVB NPC patients (7th edition of the AJCC staging system, see Supplementary Table 1 ) who were treated with IMRT at a regional cancer center between May 2017 and October 2018. All patients had complete pretreatment imaging, including MRI simulation and both contrast-enhanced and non-contrast CT scans. Before the distribution of images, CT and MR images undergo automatic rigid registration on the Monaco planning system (Version 5.1, Elekta, Sweden). Subsequently, physicians manually adjust these images with reference to bony landmarks at the skull base, such as the clivus and sphenoid sinus. This study was approved by the Institutional Review Board of Sun Yat-sen University Cancer Center (ID: B2024-111-01), and all patients provided written informed consent.

For each patient, four physicians from two different-tier cancer centers independently delineated the target volume and OARs using the Monaco planning system (Version 5.1, Elekta, Sweden). Specifically, Physicians A, B, and C, each with 5 to 10 years of experience were affiliated with a city-level cancer center. Physician D, with 6 years of experience was affiliated with a county-level cancer center. Notably, the contouring results of Physician D required collective discussion and confirmation within their department. Additionally, the delineated gold standard volumes (SVs) integrate automatic delineation algorithms with expert consensus. Initially, we utilized ABAS software (Version 2.01, Elekta AB, Stockholm, Sweden) to generate primary OARs delineations. Subsequently, three senior radiation oncology experts (each with over 10 years of experience in NPC) from a regional cancer center manually delineated the target volumes and revised OARs according to International Commission on Radiation Units and Measurements (ICRU) reports 50, 60, and 83 (18–20), which were then refined through iterative consensus until an inter-observer DSC score of over 0.90 was achieved.

The target volumes included the gross tumor volume of the nasopharynx (GTVnx), number of positive neck lymph nodes (GTVnd), high-risk clinical target volume (CTV1), and prophylactic irradiation target volume (CTV2). The volumes of the OARs encompassed the brainstem, spinal cord, lens, optic nerves, optic chiasm, pituitary, parotid glands, temporal lobes, temporomandibular joints, and mandible. Planning target volumes (PTVs) corresponding to setup uncertainties were generated for GTVnx, GTVnd, CTV1, and CTV2, as well as planning risk volumes (PRVs) were created the spinal cord, optic chiasm, optic nerves, and temporal lobe based on predefined margins.

The IOV metrics include ΔV (Equation 1), maximum-to-minimum ratio (MMR) (Equation 2), coefficient of variation (CV) (Equation 3), DSC (Equation 4), and 95% Hausdorff distance (HD95) (Equation 5) (21, 22). These are detailed as follows:

ΔV was the difference between the individual volume delineated by a physician at a county or city cancer center ( V M ) and the standard delineated volume   ( V S ):

MMR reflects the volumetric variation in organs delineated independently by different physicians.

where V m a x and V m i n are the maximum and minimum volumes of the delineated structures, respectively.

The CV indicates the dispersion of organ volume delineations among different physicians. A higher CV signifies greater IOV.

where   V s t d and V a v e represent the standard deviation and average volume of the evaluated structure, respectively.

The DSC reflects the overlap of structures delineated independently by different physicians. It is recommended that a DSC value greater than 0.7 be used as a criterion for good concordance when evaluating differences in image volume delineation (23).

where V₁ and V₂ represent the volumes delineated by two physicians, and (V₁∩V₂) is the intersecting volume of V₁ and V₂.

The HD95 reflects the similarity between the contours of two structures, defined as the maximum distance from any point on one contour to its nearest point on the other contour. The formula is:

where a and b denote the contours of structures A and B, respectively, and d(a,b) is any metric between these points.

All contoured structures were imported into the Eclipse treatment planning system (Version 15.6; Varian Medical Systems, Palo Alto, CA, USA). For each patient, the target area and OARs delineated by different physicians were independently subjected to 9-field uniform dynamic IMRT simultaneous integrated boost planning by the same experienced dosimetrist in a fully blinded manner. The same prescription and OAR dose constraints were applied to all plans with 6 MV X-ray irradiation at prescription doses of 70, 66, 60, and 54 Gy for PTVnx, PTVnd, PTV1, and PTV2, respectively, and the irradiation was delivered 33 times. The planned dose was calculated using a grid size of 2.5 mm with an anisotropic analytical algorithm (AAA). Plan optimization dose constraints were based on the 2019 international guidelines for prioritization and dose constraints for OARs in NPC (24).

To evaluate the impact of IOV on PTVs and OARs on dose distribution in treatment plans and clinical prognosis, dose prescriptions designed for individual contouring structures were mapped onto gold standardized structures. Dosimetric parameters, such as prescription dose coverage (PDC) (Equation 6) and relative dose difference ΔD_diff) (Equation 7), were used to analyze the discrepancies between each individualized and gold standard plan.

The dosing schedule for a radiotherapy program must first ensure that the planned target site is exposed to sufficient prescribed doses of radiation. Accordingly, we define the PDC of the PTV as the evaluation index of the target dose, which is calculated using the following formula:

where PTV represents the volume of the contoured PTV, and P T V 100 % is the volume of the PTV that receives 100% of the prescribed dose.

The Dmax, or D1cc for 1cc volume was used to analyze serial-type OARs such as the spinal cord and brainstem. The Dmean, or D1cc for 1cc volume was used to analyze parallel-type OARs, such as the parotid gland. ΔD_diff reflects the magnitude of dose parameter differences for OARs between plans developed at different levels of cancer centers and standard plans. It is defined as:

D M is the dose parameter from the county or city-level cancer centers, and D S is the dose parameter from the regional cancer center.

TCP, a radiobiological index for PTVnx, was computed using the Schultheiss logistic model (25), expressed as (Equation 8):

where T C D 50 denotes the dose at which no more than 50% of patients treated with radiotherapy will experience severe radiation damage 5 years posttreatment; γ50 is a unique value when TCP=0.5 and D=D50​; and EUD is the equivalent uniform dose, a measure of the dose that would produce the same radiobiological effect if the tissue or organ were uniformly irradiated, given by (Equation 9):

Following Okunieff et al. (26), the radiobiological parameter TCD50 for TCP was defined as 61.69 Gy, γ₅₀ as 3.38, and a as -8.

The normal tissue complication probability (NTCP) was assessed using a modified linear-quadratic model proposed by Zaider et al. (27), expressed as (Equations 10, 11):

where N 0 and k are nonnegative adjustable parameters that vary according to tissue or organ type; D is the dose received by normal tissue; V is the volume when the tissue is uniformly irradiated; α is the coefficient for lethal damage; and α β is the ratio of lethal to sublethal damage coefficients. The parameters required for the NTCP model calculations are provided in Supplementary Table 2 .

All data were analyzed using SPSS version 22.0 (IBM SPSS, Inc., Chicago, USA). IOV delineation discrepancies, dosimetric parameters, TCPs, and NTCPs between the four physicians and the gold standard were compared using paired t-tests or Wilcoxon signed-rank tests, with significance set at P<0.05. Spearman’s rank correlation analysis was used to assess the correlation between IOV, dosimetric parameters, and clinical prognosis parameters, with a threshold of P<0.05 indicating significant correlations, Spearman’s correlation coefficient (R) indicated the strength of correlations, and the sign of R denoted the direction of association (12).

Geometric evaluation indices of the IOV that showed significant correlations and correlation coefficients |R|>0.4 were selected and designated as predictors of delineation discrepancy risk.

A case of a 65-year-old man previously treated with IMRT and pathologically staged as T2N2M0 was randomly selected. Ten radiation oncologists from eight cancer centers independently delineated the target volumes and OARs for NPC. The structures contoured by these physicians were used in the ABAS software to establish a consensus “true structure set” by applying the STAPLE (Simultaneous Truth and Performance Level Estimation) algorithm, which serves as the gold SV within the validation cohort (28).

Based on the various delineations and planned dose distributions within the validation set, the IOV risk factors among different physicians were compared and analyzed for their correlation with clinical outcomes. This analysis aimed to validate the feasibility and generalizability of IOV risk factors as predictors of clinical prognosis.

Significant IOV was observed in the delineation of GTVnx among the four physicians. Notably, the mean GTVnx volumes were considerably higher when compared to the standard volumes (SVs). The MMR and CV for these delineations were (mean ± SD) 3.64 ± 1.60 and 0.44 ± 0.17, respectively, with average DSC values<0.6.

IOV was less for organs with clear boundaries and larger volumes, such as the brainstem, mandibles, and eyes. The average MMR was<1.8, and the CV was<0.18, with average DSC values >0.8. Conversely, IOV was considerably larger for organs with relatively obscure boundaries and smaller volumes, such as the optic nerves and pituitary. The average MMR and CV exceeded 3 and 0.5, respectively, and the average DSC values were<0.7 ( Table 1 , Figure 1 ).

Supplementary Table 3 displays the HD95 values between the delineations of the four physicians. For critical OARs, including the spinal cord, eyes, temporomandibular joint, lenses, optic nerves, and pituitary, the HD95 values for the delineations by Doctor D, a county-level physician, were markedly greater than those of the other three city cancer center physicians.

The mean PDC values for treatment plans devised at the county and city cancer centers exhibited varying degrees of decline, with the most pronounced reduction observed in one city planning group, where the average PDC% decreased by >10%. Moreover, in treatment plans originating from these centers, doses of OARs increased to differing extents. Among these organs, dose variations for the brainstem, spinal cord, and mandibles were relatively minor, with average relative dose differences<20%. In contrast, dose differences for the optic nerves and optic chiasm were considerably larger, with average relative dose differences >50% ( Table 2 ).

Figure 2 illustrates the dose distribution for plans designed by two physicians from a county or city cancer center for a specific patient. Compared to the gold standard plan Figure ( Figure 2S ), ( Figure 2A ) indicated that the plan developed at the county or city cancer center demonstrated inadequate PDC due to suboptimal target volume delineation. Conversely, Figure 2B shows an overt spillage of the prescription dose beyond the intended target volume due to an overly expansive target volume delineation by a physician from county or city cancer center.

Table 3 presents the differences in TCP and NTCP between the treatment plans designed by physicians and the gold standard plans. Compared to the gold standard plans, the TCPs for the target volumes decreased by >1% in the plans designed by the physicians, with some patients experiencing a significant decline of up to 17.83%. Additionally, the NTCPs for the physician plans increased, with the most notable increase observed in the optic chiasm, where the ΔNTCP exceeded 4.9% compared to that of the gold standard plan.

Figure 3 illustrates the correlation between geometric evaluation metrics of GTVnx, OARs, and PTVnx regarding the ΔPDC and ΔTCP. The ΔV for GTVnx strongly correlated with the ΔPDC of PTVnx (R=0.686, P<0.01). HD95 for the left mandible moderately correlated with ΔPDC of PTVnx (R=0.405, P<0.01), while other metrics demonstrated weak or no correlation with ΔPDC. Additionally, ΔV of GTVnx strongly correlated with ΔTCP of PTVnx (R=0.703, P<0.01). The ΔV of the left temporal lobe moderately correlated with ΔTCP of PTVnx (R=-0.401, P<0.01); other metrics displayed weak or no correlation with ΔTCP.

The delineation outcomes in the validation set revealed that only the ΔV of GTVnx may be a predictive factor for IOV risk. ΔV exhibited strong and moderate correlations with ΔTCP (R=0.778) and ΔPDC (R=0.596) of PTVnx, respectively ( Figures 4A, B ). In contrast, HD95 for the left mandible showed a weak or no correlation with ΔPDC of PTVnx, and the same was observed between the left temporal lobe and ΔTCP of PTVnx ( Figures 4C, D ).