Developing a segmentation cascade deep learning network based on automated prompts

Introduction

Nasopharyngeal carcinoma (NPC) is a malignancy in the region of head and neck (H&N), having a high incidence rate in endemic areas. Radiotherapy is the primary and only curative treatment for NPC (1). Particularly, intensity-modulated radiation therapy and volumetric modulated arc therapy have driven the improvement in tumor control and reduction in toxic effects in NPC survivors, providing accurate radiation delivery. Treatment planning and quality assessment for high-precision radiotherapy necessitate accurate delineation of multiple regions of interest (ROIs) by experienced radiation oncologists in accordance with recommended guidelines (e.g., RTOG 0615 Protocol). Automated segmentation (AS) approaches can largely reduce inter-practitioner variability and shorten delineation time of targets and organs at risk (OARs), thus easing the strain on radiation oncologists and reducing patient waiting time for radiotherapy (2,3).

Recently, deep-learning-based auto-segmentation (DLAS), especially the convolutional neural networks (CNNs), has demonstrated the potential as a useful tool to fulfill this need. Since Ibragimov and Xing group (4) first reported the application of DLAS to radiotherapy with CNNs, researchers (5-11) have been successfully applying DLAS to H&N computed tomography (CT) images through various networks, achieving high Dice similarity coefficient (DSC) values. However, the advancements have primarily been made in the automated outlining of OARs, while target contouring exhibits lower segmentation accuracy and offers less prominent time savings. For instance, Guo et al. (12) used 3D (three-dimensional) U-Net (3D-U-Net) and Dense-Net with combined positron emission tomography-computed tomography (PET-CT) data to reach a median DSC of 0.71 (3D U-Net) and 0.73 (Dense Net) for gross tumor volume (GTV), respectively, but with a significant decrease in the DSC for only CT data (0.32 for Dense Net) and slices with smaller tumor volumes. Huang et al. (13) obtained a DSC value of 0.785 using a 2D-U-Net while ignoring all slices with a tumor area of less than 0.5 cm², but the model’s confidence level may be overestimated because segmentation relies on information closer to the tumor’s center and the majority of these ignored slices are edge slices. Gifford et al. (14) introduced a spatially adjusted recurrent convolution U-Net (SARC U-Net), which greatly improves GTV segmentation performance at axial edges; however, with an average DSC of 0.613, which is not clinically satisfactory, the contours still require careful review and modification before treatment planning, resulting in substantial additional work. In view of this, the potential of precise automated delineation of GTV should be explored further.

Beyond ongoing network innovations, such as nnUnet (15), the two-stage segmentation method emerges as a transformative and promising solution for improving segmentation accuracy. This method, which incorporates attention mechanisms and deep learning techniques, primarily uses a CNN-based architecture to detect and localize the target region and filter out the distractors in large-size medical images before performing the precise segmentation task, with the goal of focusing processing power on the specific discerning features of the ROIs (16), making it especially suitable for small target/organ volume and complex anatomical structures. Moreover, unlike conventional single-stage networks that indiscriminately process full volumetric data, the inherent progressive refinement mechanism in such cascade frameworks achieve higher computational frugality through adaptive region focusing, which strategically allocates computing resources to anatomically relevant areas, particularly friendly to resource-constrained clinical environments.

The field of target detection is represented by region-based CNN (R-CNN) (17). Since then, single-stage neural network detection systems have also emerged, such as YOLO (You Only Look Once) (18). This method has also been clinically qualified in segmentation of different tumor sites/anatomical sites, including glioblastoma (19), brain metastases (20), esophagus (21), mediastinum (22), etc. As for H&N region, Wang et al. (23) proposed a two-stage H&N OAR segmentation framework based on 3D U-Net, which divided the segmentation into two subtasks: locating a bounding box of the OAR and segmenting the OAR from a small volume within the bounding box. The development of two-stage segmentation offers a robust solution for accurate and efficient segmentation in radiotherapy planning and treatment. However, there is no related work on target segmentation in NPC, and it is also necessary to verify whether the segmentation accuracy of the method can be further improved by training on more data.

This study proposes a two-stage DLAS cascade model for the automatic segmentation of the gross tumor volume of nasopharyngeal carcinoma (GTVnx). Built on tailored localization cues (prompts), the model integrates a self-attention mechanism and leverages the outcomes of first-stage target detection as prior knowledge. This novel network significantly enhances segmentation accuracy while simplifying training data preparation, as validated by our experimental results.

Methods

Data source

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Independent Ethics Committee of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (No. NCC2024C-1010), and informed consent was taken from all the patients.

A total of 276 patients diagnosed with NPC who received radiotherapy from 2020 to 2023 in our department were included in this study. All patients were immobilized with a thermoplastic mask (head, neck, shoulder) in the supine position. CT data were acquired via a SOMATOM Definition AS CT scanner (Siemens Healthcare, Forchheim, Germany) or a BRILLIANCE Big Bore CT scanner (Philips Healthcare, Best, the Netherlands). Contrast agent was administered to all these patients. CT images were reconstructed using a matrix size of 512×512 with a selected slice thickness of 3 mm, a tube current of 320 mAs, a tube voltage of 120 kV, a pitch of 0.9 and the pixel spacing of 0.96 mm.

All selected patients underwent MR simulation on the same day of CT simulation, using the same fixation devices and in the same position. MR images acquired were then rigidly aligned to the corresponding CT images in Pinnacle TPS (Philips Radiation Oncology Systems, Fitchburg, WI, USA) system. Radiation oncologists manually contoured GTVnx (served as ground truth) according to the international guideline (24), defining it as the primary nasopharyngeal tumor mass in planning CTs with the aid of MR image.

All GTVnx contours (primary nasopharyngeal tumor mass), served as ground truth, were delineated in accordance with the international consensus for NPC target delineation (24) by 3 radiation oncologists with ≥5 years of experience in NPC radiotherapy in planning CTs with the aid of MR image. To ensure annotation consistency and clinical authority, all initial contours underwent a two-stage validation process: (I) Inter-rater review: contours were cross-checked by another 2 senior radiation oncologists (≥10 years of experience), with discrepancies resolved via case discussion to align with protocol guidelines. (II) Multi-expert consensus: final contours were confirmed during weekly clinical rounds attended by over 20 senior professors (specializing in NPC radiation oncology) at our institution (China’s largest cancer center). This consensus process eliminated protocol-specific variations and ensured alignment with real-world clinical practice.

Proposed framework

In this study, we introduced a two-stage AS framework for nasopharyngeal carcinoma target region (GTVnx) delineation. This network adopted serially connected CNN: a localization prompt generation unit (PGU) using the deep dilated CNN (25) and a fine segmentation unit (FSU) using the deep dilated residual network (DD-ResNet) (26). The FSU used a deeper network relative to PGU. Both segmentation nets were trained independently on GTVnx and could classify every pixel in the image.

For PGU, the first-stage segmentation net, each axial CT image slice was processed through a CNN to generate a binary primary GTVnx segmentation contour. This initial segmentation served as the anatomical prior for subsequent refinement. Two distinct prompt modalities were derived from the primary segmentation contour:

• Prompt-mask: directly encoded the primary GTVnx segmentation contour.
• Prompt-box: the centroid of primary GTVnx segmentation contour was calculated and located in each CT slice. Taking the centroid point as the geometry center, a square ROI encompassing the primary GTVnx contour was selected in the CT images. “Prompt-box” is a detection box generated by the externalization of this square ROI; it was homogeneously expanded by 3 mm based on the centroid of each layer of the square ROIs. The two types of prompts served as self-attention to highlight salient regions in the way of masking the rest of the image slice (setting the pixel value outside the body region to 0).

The second-stage inputs comprised each CT image slice and prompts generated in the first stage (prompt-mask and/or prompt-box). Then the FSU network could output the fine segmentation contour on CT image slices. Figure 1 depicts the overall framework.

Figure 1 The overall framework. CT, computed tomography.

Experiment design

In order to obtain the optimal network, the performance of the two prompts mentioned and their combination was tested, respectively. Therefore, in this study, each of the four datasets for each training method was compared individually with manual contours (ground truth).

• Group A: the auto-segmentation contours of all 276 patients generated by conventional CNN model, i.e., primary GTVnx segmentation contours generated in the first stage.
• Group B: the auto-segmentation contours of all 276 patients generated by cascade network with only prompt-mask engaged in the second stage.
• Group C: the auto-segmentation contours of all 276 patients generated by cascade network with only prompt-box engaged in the second stage.
• Group D: the auto-segmentation contours of all 276 patients generated by cascade network with both prompt-mask and prompt-box engaged in the second stage.

As the available data are limited, the segmentation performance was assessed using fivefold cross-validation. Five equal-sized subgroups were randomly selected from each dataset. To ensure the reproducibility of the model, the five-fold cross-validation process was designed with strict consistency controls: all 276 cases were randomly split into 5 equal-sized subgroups to ensure identical training/test splits across all comparative experiments. This eliminated performance bias caused by variable data partitioning. The segmentation model’s parameters were “tuned” using 80% of the data (four subgroups) as the training set for each validation loop, while the remaining 20% of cases (one subgroup) served as the test set to assess the model’s performance. The procedure was repeated five times until five models are all trained and the average performance was then used as the evaluation index of the studied method. The five-fold cross-validation procedure takes full advantage of the entire datasets; all of the 2D CT slices were segmented individually. The network was trained using the stochastic gradient descent (SGD) implementation of Caffe (27).

Training process and loss function

The network was implemented in PyTorch (version 1.8) and trained on one GeForce Titan GPU with a mini-batch size of 16. We adopted the SGD optimizer with an initial learning rate of 0.01, momentum of 0.9, and weight decay of 0.0005. A polynomial learning rate decay policy was applied, reducing the learning rate by a factor of 0.1 every 50 epochs. For each fold, the network parameters are initialized by the same “he normal” (28) parameter settings. To address class imbalance between foreground (GTVnx) and background regions, we used a combination of Dice loss and cross-entropy loss as the objective function:

L=λ1∗LDice+λ2∗LCE[1]

where λ₁ =0.7 and λ₂ =0.3.

The model was trained for 200 epochs. Early stopping was implemented with a patience of 20 epochs based on validation DSC to prevent overfitting. All models converged stably within 150–180 epochs, with no significant fluctuations in validation loss (±0.02) across folds—confirming the model’s robust learning behavior and reproducibility under consistent experimental conditions. The final model was selected based on optimal performance on the validation set.

Quantitative evaluation

DSC, 95% Hausdorff distance (HD95) and mean distance to agreement (MDA) were used to quantify the segmentation accuracy of each group of data.

DSC represents the ratio of the area where AS and ground truth intersect to the total area. It ranges from 0, indicating no spatial overlap between the two segmentations, to 1, indicating the AS completely overlapped with the ground truth. DSC is calculated as (29):

DSC=2TP/(2TP+FP+FN)[2]

where TP, FP, and FN represent true positive, false positive, and false negative pixels, respectively.

The HD95 quantifies the boundary mismatch between AS contours (A) and ground truth (B). It can be ascertained using this distance with the following equation (30):

D(A,B)=a∈Amax{b∈Bmin{d(a,b)}}[3]

which implies that HD95 is the greatest of all the distances from a point in AS to the closest point in the ground truth. A smaller HD95 indicates better segmentation accuracy.

Distance to agreement (DTA) between two contours is the shortest distance from a given point on the surface of one contour to the surface of the other contour. MDA is the mean of all DTA distances (31).

The DSC, HD95 and MDA values for AS contours of GTVnx with the four automatic segmentation methods were calculated and independent-samples design was adopted to compare the performance of four methods across the three core evaluation metrics. A formal normality test was performed on the distribution of differences between each pair of datasets using SPSS was performed and the hypothesis of a normal distribution cannot be rejected. Then a unified statistical framework was applied to all three metrics: (I) Levene’s test was used to verify variance homogeneity (all P<0.001, failing to meet homogeneity); (II) Welch’s analysis of variance (ANOVA) was employed to test overall inter-group differences; (III) if significant overall differences existed, pairwise comparisons were conducted using the Games-Howell test (to control for multiple comparison errors). The statistical significance level was set at α=0.05, and analyses were performed using SPSS (version 20.0, SPSS Inc., Chicago, IL, USA). Quantitative data are demonstrated in Tables 1,2.

Results

Quantitative results

The DSC results of the AS GTVnx contours compared with ground truth in the four training groups are shown in Table 1. The DSC values in Group A (no prompts), B (with prompt-mask only), C (with prompt-box only), D (with prompt-mask and prompt-box) are 0.7489, 0.7493, 0.8162, 0.8219, respectively. The HD95 (mm) values are 12.84, 12.68, 9.28, 9.22, respectively. The MDA (mm) values are 2.23, 2.15, 1.55, 1.53, respectively.

Welch’s ANOVA revealed highly significant overall differences among the four groups across all three metrics: DSC (P<0.001), HD95 (P<0.001), and MDA (P<0.001). Further Games-Howell post-hoc tests showed (Table 2):

• For DSC, there was a highly significant increase (P<0.001) in group C vs. A, C vs. B, D vs. A, D vs. B, and D vs. C, except for group B vs. A, which did not show a significant increase (P>0.05).
• For HD95, group B vs. A showed a significant decrease (P<0.05); groups C vs. A, C vs. B, D vs. A, and D vs. B demonstrated highly significant decreases (P<0.001); and D vs. C group had no significant decrease (P>0.05).
• For MDA, all of the above groups showed a highly significant decrease (P<0.001).

These findings demonstrate three key insights:

• The Games-Howell test demonstrated that some metrics (e.g., DSC, HD95) between CNN-cascade with prompt-mask as PGU and conventional CNN were not significant, but the former still made some improvement for AS of GTVnx in terms of MDA.
• The CNN-cascade utilizing prompt-box as PGU showed substantial enhancements in GTVnx auto-segmentation accuracy, in comparison to both the CNN-cascade with prompt-mask and conventional CNN methods.
• Despite the fact that changes in HD95 values between CNN-cascade using only prompt-box and CNN-cascade combining prompt-mask and prompt-box is not that considerable, the two-channel cascade CNN model still has the highest DSC values and the lowest MDA values among all methods in this study. This indicates that the two-channel cascade CNN model exhibits the best auto-segmentation performance, with both the prompt mask and the prompt box significantly enhancing accuracy.

Qualitative results

Figure 2 illustrates the visualization of the GTVnx segmentation results for one representative patient in transverse sections. Qualitative analysis reveals performance differences among the methods, as annotated in Figure 2. Blue line shows manual contours of GTVnx (ground truth), pink line demonstrates GTVnx contours generated by method A, green line symbolizes GTVnx contours generated by method B, purple line represents GTVnx contours generated by method C and orange line displays GTVnx contours generated by method D.

Figure 2 Segmentation results for a representative patient across four methods. Blue, ground truth; pink, no-prompt CNN (Group A); green, prompt-mask (Group B); purple, prompt-box (Group C); orange, dual-prompt (Group D). CNN, convolutional neural network.

As visually highlighted in Figures 3,4, the qualitative comparison supports the quantitative findings as follows.

Figure 3 Comparative segmentation performance across four methods. The dual-prompt model (orange) achieves the best overall performance, followed by prompt-box (purple), prompt-mask (green), and no-prompt CNN (pink).

Figure 4 Poorer segmentation performance of no-prompt model and single-prompt models compared with the proposed dual-prompt model. (A) Conventional CNN (pink) yields scattered errors, absent in the dual-prompt model (orange) aligned with ground truth (blue). (B) Superior border: single-prompt models (green/purple) exhibit inward retraction, corrected by the dual-prompt model. (C) Inferior border (small GTVnx): Inward retraction in single-prompt models is more pronounced, while the dual-prompt model maintains accurate delineation. CNN, convolutional neural network; GTVnx, gross tumor volume in nasopharyngeal carcinoma.

Prompt-box vs. prompt-mask

When comparing only the purple and green lines, the purple line clearly aligns better with the ground truth. This indicates that prompt-box outperforms prompt-mask as a PGU generator for GTVnx, providing more effective spatial guidance for initial localization.

Dual-prompt superiority

In general, the auto-segmented contours with all the approaches align well with the ground truth (blue), especially for central slices. However, at the superior border, the conventional CNN model generated a few incorrect scattered points, as indicated by the red arrow in Figure 4A. The dual-prompt model shows superior alignment with the ground truth (blue), especially at the superior and inferior borders of the GTVnx, where single-prompt models exhibit inward retraction, as indicated by the red arrows in Figure 4B for the superior border and Figure 4C for the inferior border. This issue is more pronounced in slices with smaller GTVnx regions.

Visual inspection confirms the quantitative results. The dual-prompt model (orange) further refines the segmentation, bridging minor gaps between the prompt-box contour and the ground truth. This synergistic effect suggests that the prompt-mask contributes crucial local boundary information, which, when combined with the strong spatial prior of the prompt-box, yields the most anatomically faithful segmentation.

Discussion

Our study introduces a novel dual-prompt cascade framework that achieves state-of-the-art performance in GTVnx segmentation on CT-only data. The results robustly demonstrate that our two-stage, prompt-driven approach effectively addresses critical challenges in NPC target delineation, namely small lesion sensitivity and edge slice ambiguity. Qualitative analysis also shows that the dual-prompt model avoids the “inward retraction” of contours at the superior/inferior borders of GTVnx and eliminates scattered false-positive points seen in conventional CNNs. These regions are clinically critical for radiotherapy planning—under-segmentation at tumor borders may lead to insufficient radiation dose to residual tumor cells, while over-segmentation may increase toxicity to adjacent OARs (e.g., spinal cord, parotid glands). The significant performance leap over established baselines and previous methods underscores the transformative potential of integrating anatomical priors via automated prompts.

Synergistic effect of dual prompts

The core innovation of our work lies in the synergistic use of prompt-mask and prompt-box. Our study reveals that while the prompt-box alone provides a substantial performance gain (group C vs. A), the combination of both prompts yields the highest accuracy (group D). We posit that these prompts serve complementary functions: the prompt-box acts as a strong spatial attention mechanism, efficiently narrowing the processing field to a region of high probability, thereby reducing background interference and computational burden. Conversely, the prompt-mask provides a finer, shape-aware prior, guiding the FSU to refine boundaries and recapture subtle protrusions or concavities that might be lost within a rigid bounding box.

Comparison with prior work and clinical implications

Our model’s achievement of a 0.8219 DSC on CT-only data is noteworthy. It surpasses the multi-modal PET/CT approach of Guo et al. (12) (DSC 0.73), which is significant given that PET/CT is often considered the gold standard for target volume delineation in oncology. This suggests that sophisticated image analysis of CT data, guided by effective priors, can yield comparable, if not superior, accuracy without the cost and co-registration challenges of multi-modal imaging. Furthermore, we outperform the CT-based method of Huang et al. (13) (DSC 0.785) without resorting to the exclusion of smaller tumor slices, a practice that can lead to an overestimation of model performance and clinical utility.

Cascade frameworks have emerged as a promising approach for improving auto-segmentation in medical imaging. ThoraxNet (32), a 3D U-Net based two-stage cascade framework specifically designed for OAR segmentation on thoracic CT images incorporates an attention mechanism to enhance anatomical recognition of non-target structures but lacks targeted optimization for tumor regions with irregular shapes and ambiguous boundaries. Similarly, Li et al. (33) proposed a two-stage cascade deep learning model for brain metastases segmentation, which excels at small lesion detection but relies on magnetic resonance imaging (MRI)’s high soft-tissue contrast—unlike our model, which is tailored for CT-only scenarios (the most widely used modality in radiotherapy planning due to its accessibility).

Notably, most existing cascade frameworks either focus on OARs (16,23) and other tumor sites (32,33) or require multi-modal data to achieve acceptable accuracy, while few address the unique challenges of NPC GTVnx segmentation (e.g., small tumor volume, blurred edges, and proximity to critical anatomical structures). In contrast, our proposed dual-prompt cascade model fills this gap by: (I) introducing a synergistic combination of prompt-mask (boundary refinement) and prompt-box (spatial localization) to specifically tackle the edge ambiguity and small-volume sensitivity of GTVnx; (II) enabling independent training of PGU and FSU, which enhances flexibility compared to jointly trained two-stage networks and facilitates fine-tuning for clinical adaptation; (III) achieving state-of-the-art DSC (0.8219) using only CT data, outperforming multi-modal-based cascade methods and avoiding the cost and co-registration issues of multi-modal imaging. By leveraging these design choices, our model addresses the unmet need for precise, efficient, and clinically accessible GTVnx segmentation—an area where existing cascade frameworks for other anatomical sites or structures fall short.

Computational efficiency and model robustness

The proposed model decreases computing time and mitigates overfitting in deep learning-based segmentation, as validated by our experimental design and results. Specifically, two core design choices drive these advantages. For computational efficiency, the cascade architecture strategically allocates computing resources: the lightweight PGU first performs coarse GTVnx localization to filter out irrelevant background regions. The deeper FSU then focuses exclusively on the refined region-of-interest defined by prompts rather than processing the entire CT volume. This “compute-where-it-matters” approach avoids the redundant computation of single-stage networks that indiscriminately analyze full volumetric data, significantly reducing inference time while maintaining accuracy.

For robustness against overfitting, we integrated multiple safeguards: strict five-fold cross-validation with consistent data splits to ensure the model generalizes to unseen subsets, independent training of PGU and FSU to reduce mutual interference between stages, and early stopping based on validation DSC to prevent over-optimization on training data. These measures are reflected in stable convergence across all folds with minimal validation loss fluctuation and consistent performance metrics, confirming the robustness of the model.

Relevant studies support the rationale of our design. Razzak et al. (34) introduced a cascade architecture into a two-pathway-group CNN for brain tumor segmentation, demonstrating that cascade designs can improve overall performance while maintaining attractive computational complexity—aligning with our model’s efficiency-focused cascade paradigm. Similarly, Ranjbarzadeh et al. (35) proposed a cascade CNN that leverages local and global features to enforce equivariance, reducing model instabilities and overfitting. Our work builds on this insight by combining cascade design with prompt-guided attention, further enhancing robustness: the prompt-mask and prompt-box provide anatomical priors that constrain the model to clinically relevant features, reducing sensitivity to noise and irrelevant image variations. Together, these design choices and alignments with prior work confirm the model’s efficiency and robustness for clinical application.

Limitations and future directions

Despite the promising results, there are limitations in this study. First, the single-center design may limit model generalizability across institutions with different CT scanning protocols and patient demographics. However, this design was intentional for our methodological focus: the goal of this work was to first establish the superiority of the dual-prompt cascade framework in a well-controlled cohort (standardized CT acquisition, rigorous ground truth annotation) before expanding to external validation. Our single-center dataset (276 cases) is one of the largest dedicated NPC GTVnx segmentation cohorts reported to date, and the consistent performance across five-fold cross-validation provides a robust baseline for subsequent generalizability testing. For future work, we have initiated collaborations with three top cancer centers to collect multi-institutional datasets with diverse CT scanners. Detailed analyses of dosimetry impact (e.g., correlation between AS and dose coverage of GTVnx, OAR sparing) and long-term clinical outcomes (e.g., local control rate, toxicity incidence) will be conducted in our subsequent multi-center trial, where the model will be integrated into routine radiotherapy planning workflows.

Second, while our model excels on CT, the inherent lower soft-tissue contrast of CT compared to MRI remains a fundamental constraint. A compelling future direction is to develop a multi-modal cascade that integrates MRI-derived prompts to further enhance delineation in regions of ambiguous density.

Finally, the concept of “prompts” opens the door for more advanced guidance mechanisms. Exploring large language model (LLM)-driven prompts, which could incorporate structured clinical information (e.g., T-stage, tumor laterality) to generate patient-specific guidance, represents a frontier for personalized radiotherapy planning.

Conclusions

A two-stage segmentation approach can improve the precision and reliability of segmentation results, ultimately enhancing the quality of radiotherapy procedures. The prompt-driven segmentation cascade network proposed in this study has the potential for integrating into radiotherapy clinical workflow with regard to intricate target region delineation of GTVnx.

Acknowledgments

None.

Footnote

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1163/dss

Funding: This work was supported by CAMS Innovation Fund for Medical Sciences (Nos. CIFMS, 2024-I2M-C&T-A-002, 2025-I2M-XHCL-037), National Natural Science Foundation of China (Nos. 12475359, 12275357), National High Level Hospital Clinical Research Funding (Nos. 2025-LYZX-R-A01, 2022-CICAMS-80102022203), and Teaching Research Fund of Cancer Hospital of Chinese Academy of Medical Sciences (No. E2024002).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1163/coif). Y.Y. reports that this work was supported by CAMS Innovation Fund for Medical Sciences (No. CIFMS, 2025-I2M-XHCL-037). K.M. reports that this work was supported by National High Level Hospital Clinical Research Funding (No. 2025-LYZX-R-A01). X.C. reports that this work was supported by CAMS Innovation Fund for Medical Sciences (No. CIFMS, 2024-I2M-C&T-A-002), National Natural Science Foundation of China (Nos. 12475359, 12275357), National High Level Hospital Clinical Research Funding (No. 2022-CICAMS-80102022203), and Teaching Research Fund of Cancer Hospital of Chinese Academy of Medical Sciences (No. E2024002). The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Independent Ethics Committee of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (No. NCC2024C-1010), and informed consent was taken from all the patients.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Chua MLK, Wee JTS, Hui EP, Chan ATC. Nasopharyngeal carcinoma. Lancet 2016;387:1012-24. [Crossref] [PubMed]
2. Harrison K, Pullen H, Welsh C, Oktay O, Alvarez-Valle J, Jena R. Machine Learning for Auto-Segmentation in Radiotherapy Planning. Clin Oncol (R Coll Radiol) 2022;34:74-88. [Crossref] [PubMed]
3. Cardenas CE, Yang J, Anderson BM, Court LE, Brock KB. Advances in Auto-Segmentation. Semin Radiat Oncol 2019;29:185-97. [Crossref] [PubMed]
4. Ibragimov B, Xing L. Segmentation of organs-at-risks in head and neck CT images using convolutional neural networks. Med Phys 2017;44:547-57. [Crossref] [PubMed]
5. Agn M, Munck Af Rosenschöld P, Puonti O, Lundemann MJ, Mancini L, Papadaki A, Thust S, Ashburner J, Law I, Van Leemput K. A modality-adaptive method for segmenting brain tumors and organs-at-risk in radiation therapy planning. Med Image Anal 2019;54:220-37. [Crossref] [PubMed]
6. Fritscher K, Raudaschl P, Zaffino P, Spadea MF, Sharp GC, Schubert R. Deep Neural Networks for Fast Segmentation of 3D Medical Images. In: Ourselin S, Joskowicz L, Sabuncu M, Unal G, Wells W, eds. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. Lecture Notes in Computer Science. Cham: Springer; 2016:158-65.
7. Močnik D, Ibragimov B, Xing L, Strojan P, Likar B, Pernuš F, Vrtovec T. Segmentation of parotid glands from registered CT and MR images. Phys Med 2018;52:33-41. [Crossref] [PubMed]
8. Ren X, Xiang L, Nie D, Shao Y, Zhang H, Shen D, Wang Q. Interleaved 3D-CNNs for joint segmentation of small-volume structures in head and neck CT images. Med Phys 2018;45:2063-75. [Crossref] [PubMed]
9. Men K, Chen X, Zhang Y, Zhang T, Dai J, Yi J, Li Y. Deep Deconvolutional Neural Network for Target Segmentation of Nasopharyngeal Cancer in Planning Computed Tomography Images. Front Oncol 2017;7:315. [Crossref] [PubMed]
10. Ayyalusamy A, Vellaiyan S, Subramanian S, Ilamurugu A, Satpathy S, Nauman M, Katta G, Madineni A. Auto-segmentation of head and neck organs at risk in radiotherapy and its dependence on anatomic similarity. Radiat Oncol J 2019;37:134-42. [Crossref] [PubMed]
11. van Dijk LV, Van den Bosch L, Aljabar P, Peressutti D, Both S. J H M Steenbakkers R, Langendijk JA, Gooding MJ, Brouwer CL. Improving automatic delineation for head and neck organs at risk by Deep Learning Contouring. Radiother Oncol 2020;142:115-23. [Crossref] [PubMed]
12. Guo Z, Guo N, Gong K, Zhong S, Li Q. Gross tumor volume segmentation for head and neck cancer radiotherapy using deep dense multi-modality network. Phys Med Biol 2019;64:205015. [Crossref] [PubMed]
13. Huang B, Chen Z, Wu PM, Ye Y, Feng ST, Wong CO, Zheng L, Liu Y, Wang T, Li Q, Huang B. Fully Automated Delineation of Gross Tumor Volume for Head and Neck Cancer on PET-CT Using Deep Learning: A Dual-Center Study. Contrast Media Mol Imaging 2018;2018:8923028. [Crossref] [PubMed]
14. Gifford R, Jhawar SR, Krening S. Deep Learning Architecture to Improve Edge Accuracy of Auto-Contouring for Head and Neck Radiotherapy. Diagnostics (Basel) 2023.
15. Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, Heye T, Boll DT, Cyriac J, Yang S, Bach M, Segeroth M. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell 2023;5:e230024. [Crossref] [PubMed]
16. Men K, Geng H, Cheng C, Zhong H, Huang M, Fan Y, Plastaras JP, Lin A, Xiao Y. Technical Note: More accurate and efficient segmentation of organs-at-risk in radiotherapy with convolutional neural networks cascades. Med Phys 2019;46:286-92. [Crossref] [PubMed]
17. Girshick R, Donahue J, Darrell T, Malik J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. In: Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Columbus, OH, USA: IEEE; 2014:580-7.
18. Redmon J, Divvala S, Girshick R, Farhadi A. You Only Look Once: Unified, Real-Time Object Detection. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA: IEEE; 2016:779-88.
19. Tian S, Liu Y, Mao X, Xu X, He S, Jia L, Zhang W, Peng P, Wang J. A multicenter study on deep learning for glioblastoma auto-segmentation with prior knowledge in multimodal imaging. Cancer Sci 2024;115:3415-25. [Crossref] [PubMed]
20. Sadegheih Y, Merhof D. Segmentation of Brain Metastases in MRI: A Two-Stage Deep Learning Approach with Modality Impact Study. In: Rekik I, Adeli E, Park SH, Cintas C, eds. Predictive Intelligence in Medicine. PRIME 2024. Lecture Notes in Computer Science. Cham: Springer; 2024:196-206.
21. Li Z, Gan G, Guo J, Zhan W, Chen L. Accurate object localization facilitates automatic esophagus segmentation in deep learning. Radiat Oncol 2024;19:55. [Crossref] [PubMed]
22. Huang S, Han X, Fan J, Chen J, Du L, Gao W, Liu B, Chen Y, Liu X, Wang Y, Ai D, Ma G, Yang J. Anterior Mediastinal Lesion Segmentation Based on Two-Stage 3D ResUNet With Attention Gates and Lung Segmentation. Front Oncol 2020;10:618357. [Crossref] [PubMed]
23. Wang Y, Zhao L, Wang M, Song Z. Organ at risk segmentation in head and neck CT images using a two-stage segmentation framework based on 3D U-Net. IEEE Access 2019;7:144591-602.
24. Lee AW, Ng WT, Pan JJ, Poh SS, Ahn YC, AlHussain H, et al. International guideline for the delineation of the clinical target volumes (CTV) for nasopharyngeal carcinoma. Radiother Oncol 2018;126:25-36. [Crossref] [PubMed]
25. Men K, Dai J, Li Y. Automatic segmentation of the clinical target volume and organs at risk in the planning CT for rectal cancer using deep dilated convolutional neural networks. Med Phys 2017;44:6377-89. [Crossref] [PubMed]
26. Men K, Zhang T, Chen X, Chen B, Tang Y, Wang S, Li Y, Dai J. Fully automatic and robust segmentation of the clinical target volume for radiotherapy of breast cancer using big data and deep learning. Phys Med 2018;50:13-9. [Crossref] [PubMed]
27. Jia Y, Shelhamer E, Donahue J, Karayev S, Long J, Girshick RB, Guadarrama S, Darrell T. Caffe: Convolutional Architecture for Fast Feature Embedding. In: Proceedings of the 22nd ACM International Conference on Multimedia. Orlando, FL, USA: ACM; 2014:675-8.
28. He K, Zhang X, Ren S, Sun J. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification. In: 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile: IEEE; 2015:1026-34.
29. Crum WR, Camara O, Hill DL. Generalized overlap measures for evaluation and validation in medical image analysis. IEEE Trans Med Imaging 2006;25:1451-61. [Crossref] [PubMed]
30. Atallah MJ. A linear time algorithm for the Hausdorff distance between convex polygons. Inf Process Lett 1983;17:207-9.
31. Chalana V, Kim Y. A methodology for evaluation of boundary detection algorithms on medical images. IEEE Trans Med Imaging 1997;16:642-52. [Crossref] [PubMed]
32. Francis S, Jayaraj PB, Pournami PN, Thomas M, Jose AT, Binu AJ, Puzhakkal N. ThoraxNet: a 3D U-Net based two-stage framework for OAR segmentation on thoracic CT images. Phys Eng Sci Med 2022;45:189-203. [Crossref] [PubMed]
33. Li R, Guo Y, Zhao Z, Chen M, Liu X, Gong G, Wang L. MRI-based two-stage deep learning model for automatic detection and segmentation of brain metastases. Eur Radiol 2023;33:3521-31. [Crossref] [PubMed]
34. Razzak MI, Imran M, Xu G. Efficient Brain Tumor Segmentation With Multiscale Two-Pathway-Group Conventional Neural Networks. IEEE J Biomed Health Inform 2019;23:1911-9. [Crossref] [PubMed]
35. Ranjbarzadeh R, Bagherian Kasgari A, Jafarzadeh Ghoushchi S, Anari S, Naseri M, Bendechache M. Brain tumor segmentation based on deep learning and an attention mechanism using MRI multi-modalities brain images. Sci Rep 2021;11:10930. [Crossref] [PubMed]