Deep learning-based independent lymph node segmentation in esophageal cancer: a precise and efficient approach for radiotherapy planning

Introduction

Esophageal cancer (EC) is among the most commonly diagnosed malignancies and ranks as the seventh leading cause of cancer-related mortality globally (1). Radiotherapy (RT) plays a key role in the management of EC by controlling local recurrence and improving survival outcomes (2,3). However, in current clinical practice, the delineation of lymph node (LN) clinical target volumes (CTVs) in RT presents several challenges. First, it is a time-consuming and labor-intensive process (4,5). Second, although international guidelines for LNs contouring exist, manual delineation still heavily relies on individual experience and preferences, leading to significant interobserver variability (6-9).

In the context of online adaptive radiotherapy (ART), the timeliness of target contouring becomes increasingly critical. ART is effective in addressing inter-fractional variations in targets and organs at risk (OARs) by enabling daily re-contouring throughout the treatment sessions (10,11). However, for anatomically complex targets such as LN CTVs, the extended time required for manual contouring may exacerbate intra-fractional changes, as prolonged preparation can allow for additional anatomical variations to occur (12). Fortunately, the implementation of automatic segmentation that adheres to established contouring guidelines can not only improve efficiency but also enhance the quality and uniformity of contouring (13-15).

Currently, atlas-based auto-segmentation (ABAS) is a clinically available solution in the field of automatic segmentation. As the third generation of automatic image segmentation technology, it has gradually moved towards clinical application. However, this method is highly dependent on the accuracy of image matching, achieving better results only when the patient’s anatomy closely resembles that of the atlas images (16). In contrast, deep learning-based auto-segmentation (DLBAS), the fourth generation of this technology, offers superior performance in learning the characteristics of complex and low-contrast anatomy (17,18). This advantage makes DLBAS particularly promising for delineating LN CTVs. However, current research on CTV segmentation typically aims to generate a complete CTV contouring (5,19), which precludes the accurate isolation of individual LN stations within the CTV and limits the ability to select specific nodal regions as needed.

The objective of this study is to develop independent DLBAS models for distinct LN CTVs of EC. These models are designed to assist clinicians in segmenting specific LN stations and allow for flexible combinations to adapt to varying clinical stages and primary tumor locations. Furthermore, it could be used in ART to improve the efficiency of the online workflow in the future. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-2067/rc).

Methods

Patient data and structures

Contrast-enhanced computed tomography (CT) scans from 364 EC patients, all acquired within two weeks prior to surgery between October 2013 and December 2018 at Shandong Cancer Hospital and Institute, were retrospectively collected. All scans were performed using a SOMATOM Definition AS+ (Siemens Healthineers, Erlangen, Germany) scanner with the following parameters: tube voltage of 120 kVp, tube current of 200 mAs, detector configuration of 64 mm × 0.625 mm, and a beam pitch of 1.5. All enhanced CT images were exported in Digital Imaging and Communications in Medicine (DICOM) format for image feature extraction.

Two experienced radiation oncologists (each with over 20 years of experience in RT) manually delineated the regions of interest (ROIs), with precise demarcation of the boundaries of each regional LN. The target volumes were defined according to the LN classification criteria of the Japanese Esophageal Tumor Research Group, which includes the following stations: 101 (cervical paraesophageal LNs), 104 (upper thoracic paraesophageal LNs), 105 (upper thoracic paraesophageal LNs in the upper mediastinum), 106 (recurrent laryngeal nerve LNs), 107 (subcarinal LNs), 108 (middle thoracic paraesophageal LNs), 109 (main bronchus LNs), and 110 (lower thoracic paraesophageal LNs) (20). Following delineation, the data for each LN were divided into training sets and test sets for further analysis. In addition, contrast-enhanced CT scans from 16 EC patients, acquired within two weeks prior to surgery between January 2019 and December 2019, were collected for external validation. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of Shandong Cancer Hospital and Institute (project No. 2022001007). Written informed consent was obtained from all individual participants included in the study.

Network architectures

The segmentation architectures are built upon the no-new-U-Net (nnU-Net) frameworks, which consists of seven encoders and seven decoders (21), and each encoder and decoder include two three-dimensional (3D) convolutional layers. Each convolutional layer is followed by instance normalization and a Leaky rectified linear unit (Leaky ReLU) activation function. In the first layer of the network, the original data are converted to 32 channels with a feature map size of 32×256×256. To reduce model complexity, the number of channels is limited to 320 in the last three layers. In the final layer, the size of the feature map is reduced to 8×4×4.

Loss function

The loss L of the network is a combination of the dice loss L_DC and cross-entropy loss L_CE. The loss functions are shown as:

L=LDC+LCE[1]

LDC=1−2∑i=1NGiPi∑i=1NGi+∑i=1NPi[2]

LCE=−1N∑i=1NGilogPi[3]

where N is the number of voxels in each patch sample in the training set, and G_i is the i-th voxel of the ground truth (GT) contour, and P_i is the probability output of the model prediction for the i-th voxel.

Data pre-processing and post-preprocessing

The first strategy for data preprocessing is the image normalization. The normalization process typically involves scaling the intensity values to a common range, which helps in reducing variations due to different scanner settings and patient conditions. The formula for the normalization can be written as:

Inorm={I0.5,ifI≤I0.5I−I0.5I99.5−I0.5,elseifI0.5<I<I99.5I99.5,elseifI≥I99.5[4]

where I is the original intensity value, I_0.5 denotes the 0.5th percentile intensity value, and I_99.5 denotes 99.5% percentile intensity value.

Subsequently, the images for different LNs were resampled to various voxel spacings, corresponding to the median voxel spacing observed in the training set, as shown in Table S1. During the post-processing stage, the output segmentation masks were subjected to a series of operations, including the extraction of the largest connected component, erosion and dilation, and smoothing.

Training

A total of 364 patients with 1,131 LN CTVs were included for model training. The datasets for all LN stations were divided into training and test sets in a 4:1 ratio. For the segmentation network, the learning rate was 1e−3, the batch size was 2, and the patch size was 32×256×256. The optimization was conducted using stochastic gradient descent (SGD) with polynomial decay. Our training was conducted on an Intel Xeon Gold 6226R CPU and NVIDIA GeForce RTX 3090 GPU with 24GB of memory. The training process, spanning 500 epochs, was completed in 12 hours.

Evaluation

• Dice similarity coefficient (DSC): the DSC is a crucial metric for quantifying the overlap between two regions, and it is defined as:

  $$DSC\left(A,B\right)=\frac{2\left|A\cap B\right|}{\left|A\right|+\left|B\right|}$$[5] where A represents the automatic delineation by the model and B denotes the GT delineation made by human experts. Therefore, the closer the DSC value is to 1, the closer the result of auto-segmentation is to the GT delineation.

• Hausdorff distance (HD) and 95% Hausdorff distance (95HD): the HD measures the maximum distance between any two closest points on both delineations (22). The 95HD focuses on the 95th percentile of the distance distribution, indicating that 95% of the distances measured between corresponding points in the two datasets are equal to or less than this value. This approach effectively minimizes the influence of outliers on the analysis.
• Average surface distance (ASD): ASD measures the average distance between the predicted boundary and the GT boundary (23). The smaller the ASD value, the better the segmentation result.
• Clinical evaluation: the qualitative clinical evaluation was performed by two radiation oncologists using a 4-point scoring system adapted from previously established literature (24). The scoring system was defined as follows:
  • Score 0: the contour is clinically unacceptable and requires complete redrawing, with no time saved.
  • Score 1: the contour requires substantial modifications, resulting in minimal time saved.
  • Score 2: the contour requires only minor adjustments, resulting in considerable time savings.
  • Score 3: the contour is accurate and requires no corrections.
• Time evaluation: the time required for both automated segmentation with subsequent manual corrections and manual segmentation was recorded for each of the 16 patients. Besides, a Wilcoxon rank-sum test was performed using IBM SPSS version 23.0 to compare the two methods.

Results

A total of 364 patients with 1131 LNs target volumes were included for model training. The number of whole data for stations 101, 104, 105, 106, 107, 108, 109, and 110 were 150, 150, 64, 64, 301, 110, 150 and 142, respectively. The datasets for all LNs were divided into training and test sets in a 4:1 ratio. Figure 1 illustrates the flowchart for construction of DLBAS model in this study. Figure 2 shows representative images of the DLBAS-generated LN CTVs in the transverse, sagittal, and coronal planes, with different LN CTVs represented using distinct colors.

Figure 1 Flowchart for construction of DLBAS. ASD, average surface distance; CT, computed tomography; DLBAS, deep learning-based auto-segmentation; DSC, dice similarity coefficient; HD, Hausdorff distance; IN, instance normalization.

Figure 2 Representative CT images showing the comparison between DLBAS results and manual ground truth for individual lymph node stations. (A) Station 101. (B) Station 104. (C) Station 105. (D) Station 106. (E) Station 107. (F) Station 108. (G) Station 109. (H) Station 110. Red areas represent the manual delineations by radiation oncologists, green areas represent the DLBAS results, and yellow areas represent the overlap between the two. CT, computed tomography; DLBAS, deep learning-based auto-segmentation.

The quantitative metrics for different LNs on the test sets are summarized in Table 1. The mean DSC for all LNs exceeded 0.7, with stations 104, 107, and 109 demonstrating the highest overlap (mean DSC =0.79). Detailed spatial accuracy metrics, including 95HD and ASD for each station, are also presented in Table 1.

For the 16 external validation cases, the average total time for delineating single LN CTV using DLBAS followed by manual correction was 79 seconds, while the average time for manual segmentation alone was 257 seconds. The difference in time between the combined approach (DLBAS segmentation plus manual correction) and manual delineation across all ROIs was statistically significant (Table S2). The subjective evaluation of automatic segmentations was scored by two radiation oncologists, with the score distribution illustrated in Figure 3. The majority of automated segmentations received a score of ≥2, indicating they were suitable for clinical use with minimal or no modifications, accounting for 89.8% of the total cases. None of the structures were scored as “not usable”. Among the LNs, 107 and 109 showed the best performance, with 97.0% of cases scoring ≥2. In contrast, 105 showed the lowest performance, necessitating major corrections in 25.0% of cases.

Figure 3 Percentage bar chart of subjective evaluation scores for each lymph node CTV. CTV, clinical target volume.

Discussion

Long-term survival rates for EC patients remain poor, due to the high incidence of LN metastases and early recurrence (25,26). Enhancing local control rates through RT is therefore essential for the effective management of LNs (27). However, contouring LN CTVs is complex, highlighting the urgent need for tools that improve both efficiency and accuracy. To the best of our knowledge, this is the first study to train and validate DLBAS models for distinct LN CTVs of EC, offering precise and efficient segmentations that could enhance the workflow of RT.

In recent years, automatic contouring has rapidly advanced in the field of RT. Currently, automatic segmentation of OARs has gradually been incorporated into clinical practice, and extensive research has been conducted on automatic contouring of gross tumor volumes. However, among the limited studies focusing on CTVs in EC and other cancer types (5,19,28,29), most have concentrated on generating a complete CTV, which is often applicable only to specific tumor stages and clinical scenarios. For instance, the model constructed by Cao et al. was only suitable for patients with stage I or II EC who underwent radical surgery (19). Furthermore, this approach prevents the identification and verification of the LN regions included within the CTV and hinders the flexible combination of these regions according to specific clinical scenarios. In contrast, our model is designed based on individual LNs, facilitating the flexible division and combination of specific LN CTVs. This allows for the selection of target areas according to specific therapeutic objectives, including elective nodal irradiation and involved-field irradiation. Moreover, it is adaptable to different treatment strategies, including neoadjuvant, radical, and palliative RT settings.

Currently, volumetric and surface-based metrics, with the DSC being the most frequently used, serve as the primary indicators for evaluating the performance of automatic segmentation methods. Notably, achieving a high DSC for small target volumes is particularly challenging, as studies have shown that segmentation performance improves as the primary target volume increases (30). For models trained to segment individual LNs, DSC values were 0.66–0.76 for breast cancer (24), 0.71–0.91 for head and neck cancer (31), and 0.72–0.92 for lung cancer (32). For EC, model training is more difficult due to the high anatomical variability and complex LNs divisions. Nevertheless, the mean DSC for our LN CTVs still exceeded 0.7. Furthermore, our clinical evaluation results revealed that 89.8% of cases were suitable for direct clinical use with minimal or no modifications, strongly demonstrating the clinical value of this technique.

Compared with previous literature, our model demonstrated competitive quantitative performance. For instance, Wang et al. (4) reported a DSC of 0.69 for the automatic segmentation of metastatic LNs in EC. In comparison, our station-specific DLBAS model achieved higher mean DSC values ranging from 0.71 to 0.79 across different stations. Furthermore, our performance is also noteworthy when compared with recent state-of-the-art frameworks in the computer vision domain, such as the station-stratified approach by Guo et al. (33) and the transformer-based detection method by Wang et al. (34). While those studies primarily focused on anatomical LN entities, our model is specifically designed for RT CTV delineation. By incorporating clinician-driven anatomical boundaries based on the Japanese Classification of Esophageal Cancer, our model ensures that the generated contours are directly applicable to the RT planning workflow.

ART is a precise treatment approach that can adjust based on feedback during radiation therapy (11,35). In each treatment session, target contours would be generated for specific patients to optimize and evaluate the treatment plan. However, manual adjustments to complex structures and the formulation of treatment plans often consume significant time, limiting the development of ART, particularly in online ART. The introduction of artificial intelligence provides an efficient solution for ART (36-38). In our study, automatic segmentation significantly reduced contouring time on contrast-enhanced CT images, demonstrating its potential to improve efficiency in future ART applications.

Our study has the following limitations: first, our LNs segmentation for EC is based on criteria established by the Japanese Esophageal Tumor Research Group, which may require additional adjustments for clinicians accustomed to the American Joint Committee on Cancer (AJCC) classification. Second, our model is specific to EC and has not yet been validated for LNs in other cancer types. Third, all models were trained and validated using contrast-enhanced CT images in this study, which provide superior soft-tissue contrast and clearer LN visualization. However, most current online ART systems are based on cone-beam computed tomography (CBCT), magnetic resonance (MR), or non-contrast CT imaging. Therefore, the direct applicability of our model to these modalities is currently limited. The performance of the proposed model on CBCT, MR, and non-contrast CT images has not yet been evaluated and warrants further investigation. Future work will focus on multi-modality training and domain adaptation strategies to enable robust LN CTV segmentation in CBCT- and MR-guided ART workflows.

Conclusions

In conclusion, we developed and validated an automated tool for delineating LNs in the neck and chest for EC patients. Our results demonstrate that automated delineation is both efficient and accurate. The tool has the potential for implementation in routine clinical practice and provides a promising methodological foundation for future CT-guided adaptive radiation therapy applications.

Acknowledgments

During the preparation of this work, the authors used ChatGPT 4.0 to polish the language. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-2067/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-2067/dss

Funding: This study was supported by National Key Technologies Research and Development Program (grant No. 2022YFC2404605), Noncommunicable Chronic Diseases-National Science and Technology Major Project (grant Nos. 2024ZD0519900 and 2024ZD0519904), Post-Marketing Clinical Research Special Project on Innovative Drugs (grant No. WKZX2023Cx020012), Natural Science Foundation of Shandong Province (grant Nos. ZR2023LZL002, ZR2024MH007, and ZR2021LZL009), and Collaborative Academic Innovation Project of Shandong Cancer Hospital (grant No. ZF002).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-2067/coif). The 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 protocol was approved by the Ethics Committee of Shandong Cancer Hospital and Institute (project No. 2022001007). Written informed consent was obtained from all individual participants included in the study.

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. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Lin SH, Hobbs BP, Verma V, Tidwell RS, Smith GL, Lei X, et al. Randomized Phase IIB Trial of Proton Beam Therapy Versus Intensity-Modulated Radiation Therapy for Locally Advanced Esophageal Cancer. J Clin Oncol 2020;38:1569-79. [Crossref] [PubMed]
3. Minsky BD, Pajak TF, Ginsberg RJ, Pisansky TM, Martenson J, Komaki R, Okawara G, Rosenthal SA, Kelsen DP. INT 0123 (Radiation Therapy Oncology Group 94-05) phase III trial of combined-modality therapy for esophageal cancer: high-dose versus standard-dose radiation therapy. J Clin Oncol 2002;20:1167-74. [Crossref] [PubMed]
4. Wang R, Chen X, Zhang X, He P, Ma J, Cui H, Cao X, Nian Y, Xu X, Wu W, Wu Y. Automatic segmentation of esophageal cancer, metastatic lymph nodes and their adjacent structures in CTA images based on the UperNet Swin network. Cancer Med 2024;13:e70188. [Crossref] [PubMed]
5. Jin D, Guo D, Ho TY, Harrison AP, Xiao J, Tseng CK, Lu L. DeepTarget: Gross tumor and clinical target volume segmentation in esophageal cancer radiotherapy. Med Image Anal 2021;68:101909. [Crossref] [PubMed]
6. Joye I, Macq G, Vaes E, Roels S, Lambrecht M, Pelgrims A, Bussels B, Vancleef A, Stellamans K, Scalliet P, Weytjens R, Christian N, Boulanger AS, Donnay L, Van Brussel S, Moretti L, Van den Bergh L, Van Eycken E, Debucquoy A, Haustermans K. Do refined consensus guidelines improve the uniformity of clinical target volume delineation for rectal cancer? Results of a national review project. Radiother Oncol 2016;120:202-6. [Crossref] [PubMed]
7. Thomas M, Mortensen HR, Hoffmann L, Møller DS, Troost EGC, Muijs CT, Berbee M, Bütof R, Nicholas O, Radhakrishna G, Defraene G, Nafteux P, Nordsmark M, Haustermans K. Proposal for the delineation of neoadjuvant target volumes in oesophageal cancer. Radiother Oncol 2021;156:102-12. [Crossref] [PubMed]
8. Chang X, Deng W, Wang X, Zhou Z, Yang J, Guo W, et al. Interobserver variability in target volume delineation in definitive radiotherapy for thoracic esophageal cancer: a multi-center study from China. Radiat Oncol 2021;16:102. [Crossref] [PubMed]
9. Wong J, Fong A, McVicar N, Smith S, Giambattista J, Wells D, Kolbeck C, Giambattista J, Gondara L, Alexander A. Comparing deep learning-based auto-segmentation of organs at risk and clinical target volumes to expert inter-observer variability in radiotherapy planning. Radiother Oncol 2020;144:152-8. [Crossref] [PubMed]
10. Lavrova E, Garrett MD, Wang YF, Chin C, Elliston C, Savacool M, Price M, Kachnic LA, Horowitz DP. Adaptive Radiation Therapy: A Review of CT-based Techniques. Radiol Imaging Cancer 2023;5:e230011. [Crossref] [PubMed]
11. Lim-Reinders S, Keller BM, Al-Ward S, Sahgal A, Kim A. Online Adaptive Radiation Therapy. Int J Radiat Oncol Biol Phys 2017;99:994-1003. [Crossref] [PubMed]
12. Lombardo E, Dhont J, Page D, Garibaldi C, Künzel LA, Hurkmans C, Tijssen RHN, Paganelli C, Liu PZY, Keall PJ, Riboldi M, Kurz C, Landry G, Cusumano D, Fusella M, Placidi L. Real-time motion management in MRI-guided radiotherapy: Current status and AI-enabled prospects. Radiother Oncol 2024;190:109970. [Crossref] [PubMed]
13. Zhang T, Chi Y, Meldolesi E, Yan D. Automatic delineation of on-line head-and-neck computed tomography images: toward on-line adaptive radiotherapy. Int J Radiat Oncol Biol Phys 2007;68:522-30.
14. Rigaud B, Anderson BM, Yu ZH, Gobeli M, Cazoulat G, Söderberg J, Samuelsson E, Lidberg D, Ward C, Taku N, Cardenas C, Rhee DJ, Venkatesan AM, Peterson CB, Court L, Svensson S, Löfman F, Klopp AH, Brock KK. Automatic Segmentation Using Deep Learning to Enable Online Dose Optimization During Adaptive Radiation Therapy of Cervical Cancer. Int J Radiat Oncol Biol Phys 2021;109:1096-110.
15. Peroni M, Ciardo D, Spadea MF, Riboldi M, Comi S, Alterio D, Baroni G, Orecchia R. Automatic segmentation and online virtualCT in head-and-neck adaptive radiation therapy. Int J Radiat Oncol Biol Phys 2012;84:e427-33.
16. Schipaanboord B, Boukerroui D, Peressutti D, van Soest J, Lustberg T, Kadir T, Dekker A, van Elmpt W, Gooding M. Can Atlas-Based Auto-Segmentation Ever Be Perfect? Insights From Extreme Value Theory. IEEE Trans Med Imaging 2019;38:99-106.
17. 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]
18. Vrtovec T, Močnik D, Strojan P, Pernuš F, Ibragimov B. Auto-segmentation of organs at risk for head and neck radiotherapy planning: From atlas-based to deep learning methods. Med Phys 2020;47:e929-50. [Crossref] [PubMed]
19. Cao R, Pei X, Ge N, Zheng C. Clinical Target Volume Auto-Segmentation of Esophageal Cancer for Radiotherapy After Radical Surgery Based on Deep Learning. Technol Cancer Res Treat 2021;20:15330338211034284. [Crossref] [PubMed]
20. Doki Y, Tanaka K, Kawachi H, Shirakawa Y, Kitagawa Y, Toh Y, Yasuda T, Watanabe M, Kamei T, Oyama T, Seto Y, Murakami K, Arai T, Muto M, Mine S. Japanese Classification of Esophageal Cancer, 12th Edition: Part II. Esophagus 2024;21:216-69.
21. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021;18:203-11. [Crossref] [PubMed]
22. Huttenlocher DP, Klanderman GA, Rucklidge WJ. Comparing images using the Hausdorff distance. IEEE Transactions on Pattern Analysis and Machine Intelligence 1993;15:850-63.
23. Christ PF, Ettlinger F, Grün F, Elshaera MEA, Lipkova J, Schlecht S, Ahmaddy F, Tatavarty S, Bickel M, Bilic P, Rempfler M, Hofmann F, Anastasi MD, Ahmadi SA, Kaissis G, Holch J, Sommer W, Braren R, Heinemann V, Menze B. Automatic Liver and Tumor Segmentation of CT and MRI Volumes using Cascaded Fully Convolutional Neural Networks. arXiv 2017. arXiv:1702.05970.
24. Almberg SS, Lervåg C, Frengen J, Eidem M, Abramova TM, Nordstrand CS, Alsaker MD, Tøndel H, Raj SX, Wanderås AD. Training, validation, and clinical implementation of a deep-learning segmentation model for radiotherapy of loco-regional breast cancer. Radiother Oncol 2022;173:62-8.
25. Nakagawa S, Kanda T, Kosugi S, Ohashi M, Suzuki T, Hatakeyama K. Recurrence pattern of squamous cell carcinoma of the thoracic esophagus after extended radical esophagectomy with three-field lymphadenectomy. J Am Coll Surg 2004;198:205-11.
26. Shimada H, Kitabayashi H, Nabeya Y, Okazumi S, Matsubara H, Funami Y, Miyazawa Y, Shiratori T, Uno T, Itoh H, Ochiai T. Treatment response and prognosis of patients after recurrence of esophageal cancer. Surgery 2003;133:24-31.
27. Huang W, Li B, Gong H, Yu J, Sun H, Zhou T, Zhang Z, Liu X. Pattern of lymph node metastases and its implication in radiotherapeutic clinical target volume in patients with thoracic esophageal squamous cell carcinoma: A report of 1077 cases. Radiother Oncol 2010;95:229-33.
28. Hou Z, Gao S, Liu J, Yin Y, Zhang L, Han Y, Yan J, Li S. Clinical evaluation of deep learning-based automatic clinical target volume segmentation: a single-institution multi-site tumor experience. Radiol Med 2023;128:1250-61.
29. Geng J, Sui X, Du R, Feng J, Wang R, Wang M, Yao K, Chen Q, Bai L, Wang S, Li Y, Wu H, Hu X, Du Y. Localized fine-tuning and clinical evaluation of deep-learning based auto-segmentation (DLAS) model for clinical target volume (CTV) and organs-at-risk (OAR) in rectal cancer radiotherapy. Radiat Oncol 2024;19:87.
30. Schouten JPE, Noteboom S, Martens RM, Mes SW, Leemans CR, de Graaf P, Steenwijk MD. Automatic segmentation of head and neck primary tumors on MRI using a multi-view CNN. Cancer Imaging 2022;22:8. [Crossref] [PubMed]
31. van der Veen J, Willems S, Bollen H, Maes F, Nuyts S. Deep learning for elective neck delineation: More consistent and time efficient. Radiother Oncol 2020;153:180-8. [Crossref] [PubMed]
32. Shen J, Zhang F, Di M, Shen J, Wang S, Chen Q, Chen Y, Liu Z, Lian X, Ma J, Pang T, Dong T, Wang B, Guan Q, He L, Zhang Y, Liang H. Clinical target volume automatic segmentation based on lymph node stations for lung cancer with bulky lump lymph nodes. Thorac Cancer 2022;13:2897-903. [Crossref] [PubMed]
33. Guo D, Ge J, Yan K, Wang P, Zhu Z, Zheng D, Hua XS, Lu L, Ho TY, Ye X, Jin D. Thoracic Lymph Node Segmentation in CT Imaging via Lymph Node Station Stratification and Size Encoding. In: Wang L, Dou Q, Fletcher PT, Speidel S, Li S. editors. Medical Image Computing and Computer Assisted Intervention – MICCAI 2022. Springer:55-65.
34. Wang Y, Yu Q, Yan K, Li H, Guo D, Zhang L, Lu L, Shen N, Wang Q, Ding X, Ye X, Jin D. Effective Lymph Nodes Detection in CT Scans Using Location Debiased Query Selection and Contrastive Query Representation in Transformer. arXiv 2025. arXiv:2404.03819.
35. Chetty IJ, Fontenot J. Adaptive Radiation Therapy: Off-Line, On-Line, and In-Line? Int J Radiat Oncol Biol Phys 2017;99:689-91. [Crossref] [PubMed]
36. Lai J, Luo Z, Liu J, Hu H, Jiang H, Liu P, He L, Cheng W, Ren W, Wu Y, Piao JG, Wu Z. Charged Gold Nanoparticles for Target Identification-Alignment and Automatic Segmentation of CT Image-Guided Adaptive Radiotherapy in Small Hepatocellular Carcinoma. Nano Lett 2024;24:10614-23. [Crossref] [PubMed]
37. Wahid KA, Kaffey ZY, Farris DP, Humbert-Vidan L, Moreno AC, Rasmussen M, Ren J, Naser MA, Netherton TJ, Korreman S, Balakrishnan G, Fuller CD, Fuentes D, Dohopolski MJ. Artificial intelligence uncertainty quantification in radiotherapy applications - A scoping review. Radiother Oncol 2024;201:110542. [Crossref] [PubMed]
38. Teuwen J, Gouw ZAR, Sonke JJ. Artificial Intelligence for Image Registration in Radiation Oncology. Semin Radiat Oncol 2022;32:330-42. [Crossref] [PubMed]