Real-time morphological and dosimetric adaptation in nasopharyngeal carcinoma radiotherapy: insights from autosegmented fractional fan-beam CT

Abstract

Background

To quantify morphological and dosimetric variations in nasopharyngeal carcinoma (NPC) radiotherapy via autosegmented fan-beam computed tomography (FBCT) and to inform decision-making regarding appropriate objectives and optimal timing for adaptive radiotherapy (ART).

Methods

This retrospective study analyzed 23 NPC patients (681 FBCT scans) treated at Sun Yat-sen Cancer Center from August 2022 to May 2024. The inclusion criterion was as follows: ≥1 weekly FBCT via a CT-linac with ≤ 2 fractions between scans. Four deep learning-based autosegmentation models were developed to assess weekly volume, Dice similarity coefficient (DSC), and dose variations in organs at risk (OARs) and target volumes.

Results

A systematic review of autosegmentation on FBCT scans demonstrated satisfactory accuracy overall, and missegmentation was manually modified. Linear decreases in volume and/or DSC were observed in the parotid glands, submandibular glands, thyroid, spinal cord, and target volumes (R² > 0.7). The linear dose variation included coverage of the low risk planning target volume (-3.01%), the mean dose to the parotid glands (+ 2.45 Gy) and thyroid (+ 1.18 Gy), the D1% of the brainstem (+ 0.56 Gy), and the maximum dose to the spinal cord (+ 1.12 Gy). The greatest reduction in target volume coverage was noted in PGTVns, reaching 7.15%. The most significant dose changes occurred during weeks 3–6.

Conclusions

During NPC radiotherapy, the progressive dose deviations may not be corrected through repositioning alone, necessitating ART intervention. As dose variations in OARs rarely exceed 3 Gy and target coverage fluctuations remain within 10%, ART does not need to be performed frequently, and weeks 3–6 represent the most appropriate window.

Background

Intensity-modulated radiation therapy (IMRT) is the primary curative approach for nonmetastatic nasopharyngeal carcinoma (NPC). During radiotherapy, tumor regression and patient weight loss often lead to significant changes in the morphology and dosimetry of both the tumor and the organs at risk (OARs) [1,2,3,4,5]. These changes cannot be fully addressed by daily imaging-guided radiotherapy (IGRT) alone. Consequently, adaptive radiotherapy (ART) has emerged as a critical strategy to address these dynamic alterations [6,7,8,9].

However, the criteria for implementing ART in head and neck tumors, including NPC, remain poorly defined. Most studies reference factors such as changes in patient weight loss, volume of the parotid glands or target volumes [10,11,12,13], but volume changes alone may not fully capture spatial displacement. The Dice similarity coefficient (DSC), which quantifies the spatial overlap between two datasets, offers a more comprehensive assessment of morphological changes in target volumes and OARs.

Furthermore, while the parotid glands and gross tumor volumes (GTVs) were frequently studied, other OARs and target volumes were rarely mentioned [1, 2]. The morphological and dosimetric changes in these structures during radiotherapy remain poorly understood. Insufficient target volume coverage or excessive doses to OARs can compromise tumor control rates or lead to severe radiotherapy-related complications, potentially posing life-threatening risks [2, 14–15]. For example, Belshaw L et al. reported that some head and neck cancer patients experienced a 7.0–7.2 Gy increase in the spinal cord D0.1 cc (Dx% or Dy cc refers to the minimum dose received by the “hottest” x% or y cm³ of the organ) during radiotherapy [16], significantly increasing the risk of radiation-induced spinal cord injury.

To address these challenges, this study retrospectively analyzed NPC patients who received weekly fan beam computed tomography (FBCT) scans during radiotherapy. The target volumes and OARs delineated on FBCT were compared to evaluate weekly changes in volume, DSC, and dose distribution. These findings provided valuable insights for optimizing ART decision-making in NPC treatment.

Methods

Patient

The clinical and radiological data of NPC patients treated with a CT linear accelerator (uRT-linac 506c, Shanghai United Imaging Healthcare Co., Ltd. (UIH), Shanghai, China) at Sun Yat-sen Cancer Center between August 2022 and May 2024 were retrospectively collected. All patients underwent CT simulation and radiotherapy via an integrated “all-in-one” radiotherapy platform, which consolidated simulation, AI-driven autocontouring and treatment planning, quality assurance, and delivery into a unified workflow [17].

The inclusion criteria were as follows: (1) stage I–IVa NPC classified by the 8th edition of the American Joint Committee on Cancer staging system; (2) 18–65 years old, with a Karnofsky performance scale score ≥ 70; (3) definitive treatment with IMRT; (4) no previous surgery or radiotherapy for the head and neck region; (5) complete clinical data available; and (6) at least one FBCT scan performed weekly during the treatment period, with ≤ 2 radiotherapy fractions between scans. The exclusion criteria were as follows: (1) incomplete clinical data or treatment and (2) allergies to iodinated contrast agents.

Treatment planning and recalculation on fractional FBCTs

The target volumes and OARs were initially segmented via an AI-based autocontouring system, followed by manual refinement by board-certified radiation oncologists with > 10 years of NPC specialization. Planning target volumes (PTVs) were generated by applying a 3 mm isotropic margin to the GTVs and CTVs. Dose prescriptions followed institutional protocols: PGTVp (primary tumor): 70 Gy in 33 fractions; PGTVn (nodal disease): 66–70 Gy; PCTV1 (high-risk CTV): 60 Gy; PCTV2 (prophylactic CTV): 54 Gy.

Rigid registration between the planning CT and fractional FBCTs was performed. Four 3D U-Net AI models were deployed to autosegment GTVs, CTVs, and OARs on each FBCT. All AI-generated contours were subjected to physician verification and adjustment if needed. The original treatment plan was then recalculated for each FBCT. This workflow enabled longitudinal tracking of volume changes, DSC between planning CT and FBCT contours, and real-time dosimetric delivery for target volumes and OARs.

Autosegmentation of GTVs

Delineation of GTVp boundaries on noncontrast CT is challenging due to soft tissue resolution limitations. To address this, a dual-channel 3D U-Net model was implemented (Fig. 1a). Channel 1 received the FBCT image, whereas Channel 2 incorporated the rigidly registered GTVp contour from the original planning CT. This approach leverages prior spatial information to guide segmentation, achieving a DSC exceeding 0.91 in the testing cohort, demonstrating high concordance with physician-drawn contours.

Fig. 1

Networks of autosegmentation for GTVp (a), GTVns (b), CTVs (c) and OARs (d)

For lymph node delineation, two strategies were combined to improve accuracy (Fig. 1b). The first was prior spatial guidance: the GTVn contour from the planning CT was rigidly registered and mapped onto fractional FBCTs to provide an anatomical reference. The second was contextual learning: the model was trained to simultaneously predict GTVn and CTV2, enabling it to learn spatial relationships between them. This dual-task framework achieved a DSC of over 0.71.

Autosegmentation of CTVs

A prior knowledge-guided 3D U-Net architecture was developed for the delineation of CTVs (Fig. 1c). Constraint 1 was the expansion rule: CTV1 was generated by 5–10 mm isotropic expansion from GTVp. By convolving the GTVp mask with the image, a tumor-centric feature map was derived that served as the basis for the model to extract and learn multidimensional features. CTV2 was created by extending 5–10 mm beyond CTV1 and covered the cervical lymphatic drainage regions. Constraint 2 describes the spatial relationships. Specifically, (1) CTV1 was required to encompass GTVp, and CTV2 was required to encompass CTV1 and GTVn. (2) CTV1 avoided the brainstem and temporal lobes; CTV2 additionally spared the parotid glands, submandibular glands, and thyroid unless invaded. Empirical constraints were introduced by sequentially labeling CTV1, CTV2, and OARs with distinct numerical identifiers. In the testing cohort, the average DSCs for CTV1 and CTV2 were 0.92 and 0.87, respectively.

Autosegmentation of OARs

We developed a multitarget OAR delineation strategy (Fig. 1d) to address the challenges posed by the complex nasopharyngeal anatomy and diverse window width and level requirements for 35 OARs. This system categorized OARs into five anatomically related groups (Supplementary Table 1), enabling efficient spatially consistent contouring within 70 s while maintaining precision. To optimize small OAR segmentation (e.g., lenses, chiasm), we implemented a volume-adaptive dual-path model, which has been detailed in prior study, achieving an average DSC of > 0.85 across all structures, with > 0.9 for large OARs and > 0.75 for small-volume structures [18].

Loss function

The contour loss function was used to enhance the ability of the network to extract pertinent information from the prediction contours. It includes Dice loss and cross-entropy loss.

The Dice loss and cross-entropy loss were calculated as follows:

$$\:{L}_{\text{d}\text{i}\text{c}\text{e}}=\:-\frac{2}{C}\:\sum\nolimits_{k\in{C}}\frac{\sum\nolimits_{i\in{N}}G{\left(a\right)}_{k}^{i}P{\left(a\right)}_{k}^{i}}{\sum\nolimits_{i\in{N}}G{\left(a\right)}_{k}^{i}+\sum\nolimits_{i\in{N}}P{\left(a\right)}_{k}^{i}}$$

(1)

$$\:{L}_{\text{C}\text{E}}=\:1-\frac{1}{N}\:\sum\nolimits_{i\in{N}}\sum\nolimits_{k\in{C}}{G}_{i}^{k}\text{l}\text{o}\text{g}\left({P}_{i}^{k}\right)$$

(2)

where G denotes the GT and P was Pred.

The objective loss L was the weighted sum of the Dice loss L_dice and cross-entropy L_CE.

$$\:L=\:{\alpha}_{1}{L}_{\text{d}\text{i}\text{c}\text{e}}+{\alpha}_{2}{L}_{\text{C}\text{E}}$$

(3)

The coefficient weights α₁ and α₂ were empirically set to 0.5 and 0.5, respectively. A grid search was conducted to determine the values of the coefficients.

Changes in volume, DSC and dose

The volume, DSC [DSC = 2×|A∩B|/(|A|+|B|)] and dosimetric parameters for target volumes and OARs were systematically evaluated across treatment weeks 1–7. Weekly metrics were derived from FBCT scans, with week 1 data extracted from fractions 1–3 and subsequent weeks sampled at 5-fraction intervals. If no FBCT was available for a scheduled interval, the mean of adjacent fractions was used; if only one neighboring scan existed, its value was adopted.

For each structure, the average volume over weeks 1–7 was calculated and defined as Vmean. Weekly volumes were recorded as V_{week_1…,} V_{week_7}. The percentage volumetric difference of each week was calculated via the following formula: [(V_{week_x} - V_mean)/V_mean]×100% (x = 1, 2, 3… 7). To account for delineation and positioning errors, if the percentage differences of each week were < 2.5%, the structure was considered volume constant. Otherwise, a linear fit was applied to assess the trend over time. This same approach was also applied to analyze changes in the DSC.

The dose distributions were recalculated for each FBCT via the original treatment plan. The dosimetric parameters for each structure were derived from a questionnaire survey conducted by 20 experts with > 5 years of NPC specialization, with the highest number of votes being selected; the specific parameters were shown in Supplementary Table 2. The V100% (percentage coverage of 100% of the prescription dose) was selected for PGTVp and PGTVns; the V95% for PTV1 and PTV2; the D1% for the brainstem and temporallobes; the maximum dose (Dmax) for the spinal cord, optic nerves, chiasm, and temporomandibular joints (TMJs); and the D0.03 cc for the mandibles. Dmean was used for the eyes, lenses, parotid glands, submandibular glands, oral cavity, thyroid, pituitary, bony part of eustachian tubes (ETs), cochleas, internal auditory canals (IACs), larynx_supraglottic, larynx_glottis, tympanic cavities, and vestibul semises. Linear fitting was then performed.

Statistics

Origin 8 (Massachusetts, United States) was used to perform linear fitting of the volume, DSC and dosimetric parameters of the target volumes and OARs. A coefficient of determination (R²) ≥ 0.7 was considered indicative of a good linear fit.

Results

Patients

A total of 23 patients were enrolled, yielding 681 FBCT scans. The number of FBCT scans per patient ranged from 18 to 33. Specifically, 5 patients had fewer than 25 scans, while 18 patients had more than 30 scans, including 10 patients with exactly 33 scans. The clinical characteristics of all patients were summarized in Table 1.

Table 1 Patients’ characteristics

Assessment of autosegmentation accuracy

A systematic review of autosegmentation on FBCT scans demonstrated satisfactory delineation accuracy overall. However, manual corrections were required for certain structures exhibiting suboptimal autocontouring. Specifically, one FBCT scan showed incomplete delineation of the GTVp and the oral cavity; three FBCT scans from two patients exhibited an incomplete brainstem; and five FBCT scans from three patients had underestimated spinal cord contours. In addition, multiple FBCT scans revealed suboptimal contours for the chiasm, pituitary, cochlea, and vestibul semis. All remaining target volumes and OARs maintained clinically acceptable autocontours without adjustments.

Volume and DSC changes

The parotid glands, submandibular glands, thyroid, and all target volumes exhibited absolute percentage volume differences exceeding 2.5%. Linear fitting analysis revealed a gradual decrease in the parotid glands, submandibular glands, thyroid, and all target volumes, with R²>0.7 (Fig. 2a to d). The bony structures and those with absolute percentage volume differences < 2.5% were regarded as volume constants (Table 2).

Fig. 2

Volumes and DSC linear changes for OARs and target volumes

Table 2 The weekly volume and DSC of OARs and target volumes for 23 NPC patients included in this study

Structures with DSC differences < 2.5% included the brainstem, eyes, mandibles, oral cavity, temporal lobes, GTVp, and CTV1, with DSCs > 0.85 throughout treatment, and were regarded as morphologically stable. DSC changes exceeding 2.5% were observed in other structures. Linear fitting analysis revealed a decreasing trend (R² ≥ 0.7) in the parotid glands, submandibular glands, thyroid, spinal cord, and target volumes (excluding GTVp), which temporally correlated with their volumetric dynamic patterns (Fig. 2e and f). No significant trends in DSC changes were observed for the other OARs. Most of them presented DSC values < 0.7, indicating poor repeatability (Table 2).

Dosimetric changes

Dosimetric parameters for target volumes and OARs across weeks 1–7 were summarized in Table 3. The V100% for PGTVp ranged from 98.93% to 99.47%, and the V95% for PTV1 remained at 100%. However, PTV2 exhibited progressive dose degradation: the V95% decreased from 98.57% at week 1 to 94.04% at week 3, with the V100% decreasing from over 95% at weeks 1–5 to below 94% at weeks 6–7. The most pronounced target volume underdosing was observed in PGTVns, where V100% fluctuated between 85.18% and 92.61%, peaking in week 1 and reaching nadirs in weeks 3 and 5.

Table 3 The weekly dosimetric parameter value of the OARs and target volumes

Most OARs showed no specific trend in dosimetric changes. However, linear fitting models revealed significant linear trends (R²>0.7) in specific parameters (Fig. 3): Dmean to the parotid glands (+ 2.45 Gy cumulative increase, 31.80 Gy to 33.25 Gy), thyroid (+ 1.18 Gy, 45.40 Gy to 46.58 Gy), lenses, Dmax to the spinal cord, and D1% to the brainstem. Overall, all OARs maintained dose fluctuations within 3 Gy. The most significant dosimetric fluctuations in the structures with linear changes occurred during weeks 3–6. During this period, the parameter changes for the parotid glands, thyroid, spinal cord, brainstem, and PTV2 were 1.77 Gy, 0.63 Gy, 0.76 Gy, 0.41 Gy, and 1.45%, respectively.

Fig. 3

Weekly dosimetric linear changes for OARs and target volumes

Discussion

In this study, 687 sets of FBCT scans were obtained for the 23 patients, with the interval between scans not exceeding two radiotherapy fractions. Therefore, the FBCT scans were uniformly distributed and effectively captured the weekly anatomical and dosimetric variations. Our findings revealed that during NPC radiotherapy, the morphology and dosimetry of the parotid glands, thyroid and CTV2 exhibited linear changes. Despite maintaining stable volumes, the brainstem and spinal cord showed an increasing trend in dose. Overall, the fluctuations in parameters for all OARs remained < 3 Gy, and the variations in target coverage remained within 10%. The most significant morphological and dosimetric changes were observed during weeks 3–6.

Historically, radiotherapy has relied mainly on repeat simulation scans or CBCT-based evaluations to monitor morphological and dosimetric changes, yet both methods present challenges: repeat simulation scans are time-consuming, and autosegmentation in CBCT images is hindered by severe artifacts, low soft-tissue contrast, and image truncations [19]. In our study, we employed low-dose FBCT scans (integrated within the linear accelerator) to perform prefraction imaging, reducing patient time and compliance burden. Moreover, compared with CBCT, FBCT offers superior soft-tissue resolution, enabling more precise automatic delineation of target volumes and OARs.

To improve the accuracy of automatic delineation, we developed specialized algorithms for GTVp, GTVns, CTVs, and OARs, each tailored to the distinct resolutions on FBCT and the anatomical relationships between target volumes and OARs. This approach offered several advantages: first, referencing the initial treatment planning contours helped minimize the risk of omitting GTVs; second, implementing critical OAR avoidance strategies during CTV delineation better aligned with established manual practices; and third, integrating additional convolutional layers into our volume-adapted 3D U-net further refined the segmentation of small OARs, enabling the DSC for small OARs such as the cochlea and IACs to exceed 0.75 [18]. Although this approach has enhanced research efficiency and study reliability, occasional under- or mis-segmentation still necessitated manual review.

Compared with previous studies that focused primarily on volume changes in the parotid gland and target volumes [1, 20], this study examined all OARs and target volumes in the NPC. However, only the parotid glands, submandibular glands, thyroid and target volumes exhibited a linear decreasing trend. The parotid glands, submandibular glands and GTVns presented the most significant volume reductions, exceeding 20%. This finding aligned with previous studies reporting a 38.4% reduction in the GTV and a 15–35% decrease in the parotid and submandibular glands [2, 21,22,23,24]. Unexpectedly, the structures with a linear decline in the DSC closely matched those with linear volume reductions, whereas not all the structures exhibited a gradual decrease in the DSC. This finding suggested that during NPC radiotherapy, volume changes can roughly reflect morphological changes.

Morphological variations during treatment induced dynamic dosimetric fluctuations. The V100% of PTV2 decreased from 96.22% at week 1 to 93.21% at week 7, indicating a linear reduction. Similarly, the V100% of PGTVns decreased by 7% from weeks 1–4, followed by a transient increase from weeks 6–7 (Table 3; Fig. 3), potentially attributable to late-phase lymph node shrinkage enhancing target coverage. Notably, the V100% of PGTVp and the V95% of PTV1 remained relatively stable at approximately 99% and 100%, respectively, whereas Surucu et al. reported that the GTV coverage decreased from 99.3%–97.5% [3]. This discrepancy may stem from our GTVp referring to initial treatment planning contours, and only the anterior cavity was excluded, thereby minimizing anatomical alterations in GTVp and CTV1.

Progressive dose elevation was observed in the parotid glands, thyroid, brainstem, and spinal cord (Fig. 3). This finding aligned with established reports of cumulative dose escalation in the parotid gland, brainstem and spinal cord during head and neck tumor radiotherapy [3, 25], indicating that even morphologically stable structures may experience significant dosimetric drift due to subcutaneous fat reduction and superficial organ deformation. Notably, the gradual decline in target coverage and increased dose to OARs indicated that these changes were unrelated to positioning inaccuracy and could not be corrected by IGRT alone. This highlights the potential need for ART intervention. When making ART decisions, attention should be given to the morphological and dosimetric changes in the glandibular organs, brainstem, spinal cord and target volumes.

The peak dosimetric variations occurred during weeks 3–6, which aligned with findings from prior research [1, 4, 26]. During this period, deformation of the target volumes and OARs, along with dose variations, became more pronounced due to acute side effects such as pharyngeal pain and xerostomia. Nevertheless, all OAR parameter fluctuations remained <3 Gy, and target coverage variations remained ≤ 10% (Table 3). These findings suggested that frequent ART may not be necessary during NPC radiotherapy, with the optimal ART timing occurring at weeks 3–6. This finding was consistent with the proposals of Wu et al. and Gan et al. that two replans were sufficient for head‒neck IMRT [6, 27].

Conclusions

In conclusion, progressive dosimetric deviations in OARs and targets during NPC radiotherapy cannot be addressed through IGRT alone, mandating the consideration of ART. Although morphological changes may approximate dose trends in the parotid glands, thyroid, and target volumes, stable OARs such as the spinal cord and brainstem paradoxically exhibit progressive dose accumulation, requiring vigilant dosimetric surveillance. The observed limited dose fluctuations (OARs < 3 Gy, target coverage < 10%) support the avoidance of frequent ART, with weeks 3–6 representing the optimal ART window.