Automated contouring of gross tumor volume lymph nodes in lung cancer by deep learning

Automated contouring of gross tumor volume lymph nodes in lung cancer by deep learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Automated contouring of gross tumor volume lymph nodes in lung cancer by deep learning Yuling Huang, Xingxing Yuan, Longfei Xu, Junming Jian, Changfei Gong, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3999552/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 30 Sep, 2025 Read the published version in BMC Cancer → Version 1 posted 4 You are reading this latest preprint version Abstract Purpose The precise contouring of gross tumor volume lymph nodes (GTVnd) is an essential step in clinical target volume delineation. However, to the best of our knowledge, there have been no autosegmentation studies on the GTVnd in lung cancer. This study aims to propose and evaluate a deep learning model for segmenting GTVnd in lung cancer. Method Ninety computed tomography (CT) scans of patients with lymph node metastasis in lung cancer were collected, of which 75 patients were assembled into a training dataset and 15 were used in a testing dataset. A new segmentation model was constructed to enable the automatic and accurate delineation of the GTVnd in lung cancer. This model integrates a contextual cue enhancement module and an edge-guided feature enhancement decoder. The contextual cues enhancement module was used to enforce the consistency of the contextual cues encoded in the deepest feature, and the edge-guided feature enhancement decoder was used to obtain edge-aware and edge-preserving segmentation predictions. The model was quantitatively evaluated using the three-dimensional Dice Similarity Coefficient (3D DSC) and the 95th Hausdorff Distance (95HD). Results The mean 3D DSC value of the ECENet was approximately 0.72 ± 0.09, and that of the 95HD was approximately 12.65 ± 5.82 mm. The performance of ECENet was significantly improved, compared with UNet (0.46 ± 0.19 and 12.76 ± 13.12 mm, respectively). There were statistically significant differences in terms of the 3D DSC and 95HD values between from ECENet and UNet. Conclusion The proposed model could achieve the automatic delineation of the GTVnd in the thoracic region of lung cancer and showed certain advantages, making it a potential choice for the automatic delineation of the GTVnd in lung cancer, particularly for young radiation oncologists. Automatic segmentation Gross target volume lymph nodes (GTVnd) Lung cancer Deep learning Figures Figure 1 Figure 2 Figure 3 INTRODUCTION In recent years, lung cancer has become the second most common cancer with the highest mortality rate in the world [ 1 ]. Radiation therapy is an important treatment for lung cancer and the accurate contouring of the gross target volume (GTV) of the lung is essential for the precise delivery of radiotherapy [ 2 ]. Lymph Node (LN) metastasis is common in lung cancer, particularly in inoperable patients. Precision delineation of the GTV lymph nodes (GTVnd) is a necessary step in clinical target delineation for lung cancer radiotherapy and was recommended by the treatment guidelines for lung cancer [ 3 – 4 ]. Although positron emission tomography CT (PET-CT) is the most appropriate method by which to identify the GTVnd, not all lung cancer patients undergo PET-CT scanning because of economic and other factors, particularly in underdeveloped regions [ 5 – 6 ]. Therefore, it is mainly segmented manually by oncologists in clinical application, highly dependent on manual experience, and results in significant heterogeneity, despite oncologists following the same guidelines [ 7 – 8 ]. Moreover, it is particularly difficult for beginners or inexperienced clinical oncologists to delineate the GTVnd. Therefore, it is of great clinical value to achieve automatic and accurate segmentation of the GTVnd in patients with lung cancer with LNs metastasis. With the development of deep learning, convolutional neural network (CNN)-based methods have been successfully applied to the automatic contouring of lung cancer. Significant progress has been made in the study of automatic organ-at-risk (OAR) segmentation [ 9 – 12 ], the GTV of primary tumors [ 13 ], and the clinical target volume (CTV) in postoperative radiation therapy (PORT) [ 14 – 15 ]. Ibragimov et al. used CNNs for OARs segmentation in H&N CT images and obtained Dice Similarity Coefficient (DSC) values that varied from 37.4% for the chiasm to 89.5% for the mandible [ 16 ]; however, no target was segmented. UNet, proposed by Ronneberger et al. is primarily used for the segmentation of biomedical images [ 17 ] and did not specialize in target segmentation in the field of radiotherapy. Recently, Shen et al. proposed the DiUNet model to automatically delineate the CTV of patients with lung cancer with LN metastasis [ 18 ]. Although the mean value of the 3D DSC was greater than 0.7, the GTVnd was manually contoured and added to the CT slice as a second input in the CNN architecture. To the best of our knowledge, no studies have been conducted on the automatic segmentation of the GTVnd in lung cancer. The autodelineation of the GTVnd is difficult because of its unclear boundaries and variability in the size and shape. Although UNet, which was based on encoder-decoder architecture, has shown effective performance in conducting multilevel feature fusions, the network is insufficient to represent high-level features, particularly structures that are of significant importance for GTVnd recognition. Therefore, we propose a novel edge-guided contextual cue enhancement network (ECENet) to improve the segmentation performance of the GTVnd by introducing a contextual cue module and incorporating edge guidance into the feature learning of the segmentation network. A contextual cue modeling module is implemented to enhance spatial and semantic information at a distance. To the best of our knowledge, this is the first study to attempt to use this deep learning approach to delineate the GTVnd in lung cancer, hoping to improve the efficiency of clinical work and accuracy of target delineation. METHODS Patient Datasets CT data from 90 patients with small cell lung cancer (SCLC) with involved thoracic regional LN were collected from January 2021 to January 2024. The number of slices per patient ranged from 70 to 105, with a total of 8,758 slices. The data were 512 × 512, with a pixel pitch of 1.27 × 1.27 mm and a thickness of 5 mm, and they were acquired using a Brilliance CT Big Bore (Siemens Healthcare). The private information of the patients was kept confidential during data collection and processing. GTVnd was defined as the area of the LN stations with a maximum short axis of > 1 cm. The GTVnd was delineated in the mediastinal and lung windows according to Radiation Therapy Oncology Group (RTOG) guidelines [ 19 ]. GTVnd was manually delineated by senior radiation oncologists. Architecture of ECENet As shown in Fig. 1 , based on the encoder-decoder architecture of UNet [ 20 ], the encoder of the proposed ECENet extracts the multiscale features from the input image, including low-level spatial details and high-level semantic cues. The proposed contextual cue enhancement module is then used to enforce the consistency of the contextual cues encoded in the deepest feature. The proposed ECENet decoder aims to combine the extracted multiscale features. In this process, for the three lowest layers, the mask of the GTVnd and the edge of the GTVnd area are used to supervise and guide the simultaneous learning of segmentation mask prediction and edge prediction. In this manner, using the final learned edge-aware features, an edge-preserving segmentation prediction can be obtained. (1) Contextual cues enhancement module The proposed contextual cue enhancement module consists of convolution kernel prediction and contextual cue aggregation submodules. As shown in Fig. 2 , the convolution kernel prediction submodule is first used to model both the long-distance and short-distance semantic and spatial relations between any two pixels in the manner of matrix multiplication, thereby generating a spatial-adaptive contextual convolution kernel. The generated contextual convolution kernel is then applied to the original feature, leading to aggregated contextual cue enhanced features. Concretely, for the deepest feature \(\text{X}\in {\text{R}}^{\text{H}\times \text{W}\times \text{C}}\) , the \(\text{X}\) is first transformed with two independent convolutional operations, generating two features: the key \(\text{K}\in {\text{R}}^{\text{H}\times \text{W}\times \text{C}}\) and the query \(\text{Q}\in {\text{R}}^{\text{H}\times \text{W}\times {\text{s}}^{2}}\) , where H, W and C denote the height, width and number of channels of the features, respectively, and s denotes the scale of the contextual convolution kernel. In order to model the relation between any two pixels, the feature of any pixel locations should be interacted in an explicit manner. To this end, the features of key and query are reshaped into a 2D shape: \({K}^{{\prime }}\in {\text{R}}^{(\text{H}\times \text{W})\times \text{C}}\) , \({Q}^{{\prime }}\in {\text{R}}^{（\text{H}\times \text{W}）\times {\text{s}}^{2}}\) . For the key feature, the length of the feature vector for every pixel location is C. For the query feature, the length of the feature vector for every pixel location is \({\text{s}}^{2}\) . To explicitly model the interaction between the feature vector of every pixel location of the key feature and the query feature, the element-wise matrix multiplication between the two feature vectors is calculated: $${\text{S}}^{{\prime }}\left(\text{i},\text{j}\right)=\sum _{\text{q}=1}^{\text{H}\times \text{W}}{\text{Q}}^{{\prime }}\left(\text{q},\text{i}\right)\times \text{K}{\prime }(\text{q},\text{j})$$ 1 where \(\text{i}=\text{1,2},\dots ,{\text{s}}^{2}\) and \(\text{j}=\text{1,2},\dots ,\text{C}\) . Thus, the contextual cues that enhance feature \({\text{S}}^{{\prime }}\) can be obtained. This operation can be rewritten as follows: $${\text{S}}^{{\prime }}={\text{Q}{\prime }}^{\text{T}}\times \text{K}{\prime }$$ 2 where \({\text{S}}^{{\prime }}\in {\text{R}}^{{\text{s}}^{2}\times \text{C}}\) and \({\text{Q}{\prime }}^{\text{T}}\) is the transpose of \(\text{Q}{\prime }\) . Then, in order to obtain the final contextual convolution kernel, the \({\text{S}}^{{\prime }}\in {\text{R}}^{{\text{s}}^{2}\times \text{C}}\) should be reshaped into \(\text{S}\in {\text{R}}^{\text{s}\times \text{s}\times \text{C}}\) . Furthermore, to generate the spatially adaptive weight factor \(\text{M}\) , the contextual convolution kernel \(\text{S}\) and the original input feature \(\text{X}\) were calculated using depth-wise convolution. First, \(\text{S}\in {\text{R}}^{\text{s}\times \text{s}\times \text{C}}\) is splits into C kernels, with each kernel being \(\text{s}\times \text{s}\) in shape. Then, these kernels are applied for every feature channel of the input feature \(\text{X}\in {\text{R}}^{\text{H}\times \text{W}\times \text{C}}\) . Finally, the weight factor \(\text{M}\) is multiplied with the original input feature \(\text{X}\) in an element-wise manner to generate the final contextual cue enhanced features. (2) Edge-guided feature enhancement The decoder is designed to effectively combine the multiscale features generated by the encoder in a layer-wise manner. In this process, for the lowest three layers, both the mask of the GTVnd region and the edge of the GTVnd region are used for supervisions. The features of the decoder learn to predict the segmentation mask and edge of the GTVnd simultaneously. In this manner, the learned edge-aware features can generate edge-preserving predictions. To obtain the ground truth of the edge of the GTVnd, we use the following operations to transform the mask of the GTVnd into the edge of the GTVnd. For each pixel, if the ground truth label of this pixel is different from that of any one of the eight neighboring pixels, then this pixel is an edge pixel. Experiments Complete CT image data were randomly divided into a training dataset containing 75 cases and a testing dataset containing 15 cases in a ratio of 5:1. ECENet was trained for GTVnd contouring with the 75 cases in the training dataset and it was applied to predict 15 cases in the testing dataset. Stochastic gradient descent was used in the optimization, and the initial learning rate was set to 0.001. An NVIDIA RTX 4090 GPU was used in the experiments. Because ECENet was developed based on UNet, the performance of this model in contouring the GTVnd in lung cancer was comprehensively evaluated by comparing its performance with that of UNet. UNet was trained separately for GTVnd contouring under the same settings as ECENet. Quantitative Evaluation Metrics To quantify the contouring accuracy, the Dice Similarity Coefficient (DSC) [ 21 ] and 95th percentile Hausdorff Distance (95HD) [ 22 ] were used. The DSC is defined as follows: $$DSC{\kern 1pt} {\kern 1pt} {\kern 1pt} \left( {A\invalidcharacter B} \right){\text{=}}\frac{{2\left| {A \cap \left. B \right|} \right.}}{{\left| A \right|{\text{+}}\left| B \right|}}$$ 3 where A represents the predicted mask, B the GT mask, and | A ∩ B | the intersection of A and B . The DSC measured the relative volumetric overlap between two segmented masks, with a higher value indicating a higher overlap ratio and a value of one meaning that the two masks are identical. The 95 HD is defined as: $$\begin{gathered} 95HD\left( {A,B} \right){\text{=percentile}}\left[ {{\text{h}}\left( {A,B} \right) \cup {\text{h}}\left( {B,A} \right),95th} \right] \hfill \\ h\left( {A,B} \right)=\mathop {\hbox{max} }\limits_{{a \in A}} \mathop {\hbox{min} }\limits_{{b \in B}} \left\| {a - b} \right\| \hfill \\ h\left( {B,A} \right)=\mathop {\hbox{max} }\limits_{{b \in B}} \mathop {\hbox{min} }\limits_{{a \in A}} \left\| {b - a} \right\| \hfill \\ \end{gathered}$$ 4 where \(\left\| . \right\|\) represents the Euclidean norm of points A and B . A = {a 1 , a 2 , …, a n1 } and B = {b 1 , b 2 , …, b n2 } represent two finite point sets. The 95HD value indicates the 95th percentile of mismatches between A and B . The 95HD reflected the alignment between the wo contours, with a higher value indicating a larger difference. The DSC and 95HD were calculated at the 3D level, and the mean standard deviation (SD) values were calculated by averaging the values. RESULTS Figure 3 shown the GTVnd segmentation slice of one patient, including the transversal, coronal, and sagittal sections. The GTVnd was located in zones 2 and 4 of the lung cancer LN station, with a volume of 137.22 cc. The 3D DSC of this case from UNet and ECENet were 0.40 and 0.83, respectively, and the 95HD values were 6.93 mm, and 8.54 mm, respectively. Compared to the UNet models, ECENet performed much better in the segmentation of the GTVnd, which was highly consistent with the oncologist's manual segmentation contours. The 3D DSC and 95HD were calculated at a 3D level for the testing dataset and the mean ± standard deviation (SD) values for UNet and ECENet were calculated by averaging the values. The quantitative evaluation results were summarized in Table 1 . The mean 3D DSCs for UNet and ECENet were 0.46 ± 0.17 and 0.72 ± 0.09, respectively, and the mean 95HDs were 12.76 ± 13.12 mm and 12.65 ± 5.82 mm, respectively. The proposed ECENet tended to achieve higher accuracy than did UNet by achieving greater 3D DSC values and lower 95HD values. Moreover, the proposed model performed more robustly on the testing dataset, displaying a small average standard deviation. There was a statistically significant difference between the different models for both 3D DSC (P < 0.001) and 95HD (P < 0.001). In addition, the mean volume of GTVnd was 46.06 ± 42.33 cc. The standard deviation values are large, owing to the large differences in the GTVnd volume for each case in the testing dataset. Table 1 The statistics of 3D DSC and 95HD with the volume of the GTVnd for UNet and ECENet in the testing dataset. DSC ± SD 95HD ± SD (mm) UNet 0.46 ± 0.19 12.76 ± 13.12 ECENet 0.72 ± 0.09 12.65 ± 5.82 DISCUSSION Few studies have been conducted on the automatic delineation of the GTVnd for LN metastasis in lung cancer, and oncologists mainly perform manual contours based on their clinical experience, which leads to heterogeneity between different oncologists and intraobservers at different times [ 23 , 24 ]. However, consistent target delineation is essential for improving radiotherapy outcomes. Since the artificial intelligence successfully performed automatic target delineation, particularly in improving the accuracy of contours and reducing heterogeneities [ 25 ], an ECENet based on UNet was constructed to automatically delineate the contours of the GTVnd in lung cancer accurately and consistently. The segmentation performance of the model was also evaluated. The DSC value was commonly used to assess the accuracy of autosegmentation during radiation treatment planning. Because ECENet was developed based on UNet, its performance in contouring the GTVnd in lung cancer was comprehensively evaluated by comparing it with UNet. The experimental results in Table 1 demonstrated that the proposed outperforms UNet models. The DSC values increased from 0.46 to 0.72, revealing that it is feasible and effective to develop ECENet for segmenting GTVnd contours in lung cancer by implementing a contextual cue enhancement module and edge-guided feature enhancement. Moreover, for the segmentation of the GTVnd, the DSC values have been reported to be 0.46 [ 26 ] and 0.62 [ 27 ], whereas our proposed model demonstrated a high DSC value of 0.72. These models were not tested using the same dataset; therefore, it would be unfair to say that our proposed model was superior. However, it was reasonable to conclude that ECENet yields good results. Although the segmentation accuracy of GTVnd is better than the previously reported, it still displays inferior performance compared with the published GTV or OAR contouring models [ 28 – 30 ], in which the mean DSC values tend to be greater than 0.8. This reason was mainly because the GTV and OAR were characterized differently from the GTVnd. While most GTV or OAR regions had well-defined boundaries, the GTVnd did not, and the contour boundaries of GTVnd could easily be confused with the various complex anatomical structures around it. Moreover, there were considerable differences in the shape, volume, and location of the GTVnd among patients. As a result, the more complex GTVnd contouring performed inferiorly compared to the GTV or OAR contouring models. This was a single-center study with a small sample size, and the segmentation accuracy of ECENet for GTVnd was inferior compared to that of the reported GTV or OAR contouring. However, this model is superior to UNet and shows a higher DSC value than the reported model for the segmentation of the GTVnd. Since Zijdenbos et al. suggested that a DSC value of > 0.7 represents good overlap [ 31 ], this model with a DSC value of 0.72 can accurately achieve the automatic contouring of GTVnd in lung cancer. However, the segmentation accuracy of the GTVnd remains unsatisfactory for clinical application. Moreover, the dose deviation due to imperfect GTVnd segmentation needs to be further calculated, and its impact on treatment efficacy should be evaluated to avoid the under dosage of target volumes or an overdose being delivered to normal tissues. In general, this model can provide a potential benefit for the automatic contouring of GTVnd and CTV in lung cancer. In the future, the dataset size could be expanded to further optimize the model and improve its applicability and overall accuracy. With further revision of the dose deviation due to imperfect GTVnd segmentation in treatment planning, the model will be abled able to be clinically applied. Simultaneously, this technology can be explored for the automatic segmentation of the GTVnd in other types of tumors. CONCLUSION The automatic and accurate segmentation of the GTVnd is key to achieving the automatic delineation of the CTV for lung cancer with LN metastasis. In this study ECENet was proposed for the fully automatic and accurate contouring of GTVnd in lung cancer. The model was based on UNet with edge guidance and spatial context enhancement. The results of the quantitative evaluation this model showed that it could accurately delineate the contours of the GTVnd. However, a multicenter clinical evaluation with more cases is needed to fully apply the model in clinical practice. In conclusion, this study explored the application of deep learning to the automatic delineation of the GTVnd in lung cancer. ECENet has shown the potential to improve the consistency of contouring, which also provides the potential to positively contribute to the automatic delineation of the GTVnd. Declarations Funding This work was supported by Jiangxi Cancer Hospital Doctoral Research Start-up Fund (No. BSQDJ202301) and the Jiangxi Cancer Hospital Scientific Research Open Fund Project (No. KFJJ2023YB19). Data availability All data generated or analyzed during this study are included in this published article and its supplementary information files. Competing Interest The authors declare no competing interests. Ethics approval and consent to participate This study was approved by the ethics committee of the Jiangxi Cancer Hospital (2023ky192). Informed consent was obtained from all patients included in this study. References Sung H, Ferlay J, Siegel R L, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA: A Cancer Journal for Clinicians, 2021, 71(3): 209-249. Van Rikxoort E M, de Hoop B, Viergever M A, et al. Automatic lung segmentation from thoracic computed tomography scans using a hybrid approach with error detection[J]. Medical physics, 2009, 36(7):2934–2947. Ettinger DS, Wood DE, Akerley W, Bazhenova LA, Borghaei H, Camidge DR, et al. Non-small cell lung cancer, version 6.2015. J National Comprehensive Cancer Network.2015;13(5):515–24. De Ruysscher D, Faivre-Finn C, Nestle U, Hurkmans CW, Le Péchoux C, Price A, et al. European organisation for research and treatment of cancer recommendations for planning and delivery of high-dose, high-precision radiotherapy for lung cancer. J Clin Oncol. 2010;28(36):5301–10. Petersen H, Holdgaard PC, Madsen PH et al (2016) FDG PET/CT in cancer: comparison of actual use with literature-based recommendations. Eur J Nucl Med Mol Imaging 43:695–706. Fletcher JW, Djulbegovic B, Soares HP et al (2008) Recommendations on the use of 18F-FDG PET in oncology. J Nucl Med 49:480–508. Benjamin E. Nelms, Wolfgang A. Tomé, Greg Robinson, James Wheeler. Variations in the Contouring of Organs at Risk: Test Case From a Patient With Oropharyngeal Cancer[J]. International Journal of Radiation Oncology, Biology, Physics, 2010, 82(1):368-378. Coen Rasch, Avraham Eisbruch, et al. Irradiation of paranasal sinus tumors, a delineation and dose comparison study[J]. International Journal of Radiation Oncology, Biology, Physics, 2002, 52(1):120-7. Uzelaltinbulat S, Ugur B. Lung tumor segmentation algorithm[J]. Procedia Computer Science, 2017, 120: 140-147. Soltani-nabipour J, Khorshidi A, Noorian B. Lung tumor segmentation using improved region growing algorithm[J]. Nuclear Engineering and Technology, 2020, 52(10): 2313-2319. Yang B L, Xiang D H, Yu F H, et al. Lung tumor segmentation based on the multi-scale template matching and region growing[C]//Proc. SPIE 10578, Medical Imaging 2018: Biomedical Applications in Molecular, Structural, and Functional Imaging, 2018: 105782Q. Ballangan C, Wang X, Fulham M, et al. Lung tumor segmentation in PET images using graph cuts[J]. Computer Methods and Programs in Biomedicine, 2013, 109(3): 260-268. Mi H M, Petitjean C, Dubray B, et al. Automatic lung tumor segmentation on PET images based on random walks and tumor growth model[C]//IEEE International Symposium on Biomedical Imaging, 2014: 1385-1388. Nian Y J, Li M, Cui H R, et al. Graph-based unsupervised segmentation for lung tumor CT images[C]//IEEE International Conference on Computer and Communications, 2017: 1884- 1888. Oda H, Bhatia K K, Roth H R, et al. Dense Volumetric Detection and Segmentation of Mediastinal Lymph Nodes in Chest CT Images[C]. MORI K, PETRICK N, eds. Medical Imaging 2018: Computer Aided Diagnosis. SPIE, 2018: 1. Ibragimov B, Xing L. Segmentation of organs-at-risks in head and neck CT images using convolutional neural networks. Med Phys (2017) 44(2):547–57. doi:10.1002/mp.12045. Ronneberger O, Fischer P, Brox T. UNet: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical Image Computing and Computer Assisted Intervention, 2015, 9351: 234-241. Jie Shen, Fuquan Zhang, Mingyi Di, et al. Clinical target volume automatic segmentation based on lymph node stations for lung cancer with bulky lump lymph nodes. Thorac Cancer. 2022 Oct; 13(20): 2897–2903. Bradley JD, Paulus R, Komaki R, et al. Standard-dose versus high-dose conformal radiotherapy with concurrent and consolidation carboplatin plus paclitaxel with or without cetuximab for patients with stage IIIA or IIIB non-small-cell lung cancer (RTOG 0617): a randomised, two-by-two factorial phase 3 study. Lancet Oncol. 2015;16(2):187–99. Ronneberger O, Fischer P, Brox T. UNet: Convolutional networks for biomedical image segmentation. arXiv preprint arXiv:1505.04597, 2015 Dice LR. Measures of the amount of ecologic association between species. Ecology. 1945;26(3):297–302. Huttenlocher DP, Klanderman GA, Rucklidge WJ. Comparing images using the Hausdorff distance. IEEE Trans Pattern Anal Mach Intell. 1993;15(9):850–63. Mercieca S, Belderbos JSA, van Baardwijk A, et al. The impact of training and professional collaboration on the interobserver variation of lung cancer delineations: a multi-institutional study. Acta Oncol. 2019;58(2):200–8. Guangying Z, Tingyi X, Lvhua W, et al. Consensus and controversies on delineation of radiotherapy target volume for patients with non-small cell lung cancer. Chin J Rasiat Oncol. 2008;17(6):432–6. Barkati M, Simard D, Taussky D, et al. Magnetic resonance imaging for prostate bed radiotherapy planning: An inter- and intra-observer variability study. J Med Imaging Radiat Oncol (2016) 60:255–9. doi: 10.1111/1754-9485.12416. Yang J, Beadle BM, Garden AS, et al. Auto-segmentation of low-risk clinical target volume for head and neck radiation therapy. Pract Radiat Oncol (2014) 4(1): e31–7. doi:10.1016/j.prro.2013.03.003 Kuo Men, Xinyuan Chen, Ye Zhang, et al. Deep Deconvolutional Neural Network for Target Segmentation of Nasopharyngeal Cancer in Planning Computed Tomography Images. Front Oncol. 2017 Dec 20;7:315. doi: 10.3389/fonc.2017.00315. PMID: 29376025; PMCID: PMC5770734. Thibault Marin, Yue Zhuo, Rita Maria Lahoud, et al. Deep learning-based GTV contouring modeling inter- and intra- observer variability in sarcomas, Radiotherapy and Oncology, Volume 167, 2022, Pages 269-276, ISSN 0167-8140, https://doi.org/10.1016/j.radonc.2021.09.034. Lang, D.M., Peeken, J.C., Combs, S.E.,et al. Deep Learning Based GTV Delineation and Progression Free Survival Risk Score Prediction for Head and Neck Cancer Patients. Lecture Notes in Computer Science, vol 13209. Springer, Cham. https://doi.org/10.1007/978-3-030-98253-9_14. Shi, F., Hu, W., Wu, J. et al. Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. Nat Commun 13, 6566 (2022). https://doi.org/10.1038/s41467-022-34257-x. Zijdenbos AP, Dawant BM, Margolin RA, Palmer AC. Morphometric analysis of white matter lesions in MR images: method and validation. IEEE Trans Med Imaging (1994) 13(4):716–24. doi:10.1109/42.363096 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 30 Sep, 2025 Read the published version in BMC Cancer → Version 1 posted Editorial decision: Revision requested 15 Mar, 2024 Submission checks completed at journal 15 Mar, 2024 Editor assigned by journal 15 Mar, 2024 First submitted to journal 29 Feb, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3999552","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":279877096,"identity":"d0036655-8670-4f47-b8b1-2d78a7f7c3d7","order_by":0,"name":"Yuling Huang","email":"","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":false,"prefix":"","firstName":"Yuling","middleName":"","lastName":"Huang","suffix":""},{"id":279877098,"identity":"9378922a-5b55-4798-bec1-57b35d0415bc","order_by":1,"name":"Xingxing Yuan","email":"","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":false,"prefix":"","firstName":"Xingxing","middleName":"","lastName":"Yuan","suffix":""},{"id":279877099,"identity":"011806a9-f6fc-4976-ae8b-b2ff98c20b86","order_by":2,"name":"Longfei Xu","email":"","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":false,"prefix":"","firstName":"Longfei","middleName":"","lastName":"Xu","suffix":""},{"id":279877100,"identity":"922185e1-c874-4b11-a79b-5405d4a61167","order_by":3,"name":"Junming Jian","email":"","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":false,"prefix":"","firstName":"Junming","middleName":"","lastName":"Jian","suffix":""},{"id":279877102,"identity":"1cfeaed7-f357-425e-b116-d7e53ee183ba","order_by":4,"name":"Changfei Gong","email":"","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":false,"prefix":"","firstName":"Changfei","middleName":"","lastName":"Gong","suffix":""},{"id":279877104,"identity":"2304401a-6a38-431d-b338-d7db50a1254e","order_by":5,"name":"Yun Zhang","email":"","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":false,"prefix":"","firstName":"Yun","middleName":"","lastName":"Zhang","suffix":""},{"id":279877106,"identity":"7d844ae5-1cf8-4434-8958-5e9d150943e6","order_by":6,"name":"Wenheng Zheng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYDACCTB5gIefvfEBiMXDR6wWOcmewwZgLWzEajE2mJEM1sJAUIv87OZjD7+23UncIPmY8TFvjp0MGwPzw0c38GhhnHMs3Vi27VnidulkZsOZ25KBDmMzNs7Bo4VZIsdMWnLb4cSds/OPSXzcxgzUwsMmjU8Lm0T+N7CWDTcPs0kkbqsnrIVHIodN8uO2w8YGN5jZgLYcJqxFQiLNTJrx32FgIIP9cpyHjZmAX+RnJD+T/HHmMDAqDwNDbFu1PT9788PH+LSAADMPKpeAchBg/EGEolEwCkbBKBjBAABv9EVYepaP3AAAAABJRU5ErkJggg==","orcid":"","institution":"Department of Radiation Oncology, Jiangxi Cancer Hospital of Nanchang University, Nanchang, Jiangxi","correspondingAuthor":true,"prefix":"","firstName":"Wenheng","middleName":"","lastName":"Zheng","suffix":""}],"badges":[],"createdAt":"2024-02-29 11:47:58","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3999552/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3999552/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12885-025-14794-6","type":"published","date":"2025-09-30T15:58:17+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":53018026,"identity":"4aaa3517-e223-415c-bdc6-af20174e9f00","added_by":"auto","created_at":"2024-03-19 16:14:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":79662,"visible":true,"origin":"","legend":"\u003cp\u003eOverall network structure of the ECENet.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-3999552/v1/19c7322cf38c9bba255b4562.png"},{"id":53019426,"identity":"ed1ecf3f-70b3-43db-90a4-db118db7a9cb","added_by":"auto","created_at":"2024-03-19 16:22:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":78325,"visible":true,"origin":"","legend":"\u003cp\u003eDetails of the proposed contextual cues enhancement module.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-3999552/v1/0ceacdae56e9584e3ddb4c07.png"},{"id":53018028,"identity":"41f976b5-2e07-4a7e-ba5b-87c7b677671b","added_by":"auto","created_at":"2024-03-19 16:14:40","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":402429,"visible":true,"origin":"","legend":"\u003cp\u003eExample of GTVnd segmentation slice for a patient predicted from UNet and ECENet. (A) GTVnd on three transversal planes, (B) GTVnd on a coronal plane, (C) GTVnd on a sagittal plane.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-3999552/v1/a4c2fc0867cb97431f245aba.png"},{"id":92884655,"identity":"082cc773-16ad-4ba7-822c-858416a0f070","added_by":"auto","created_at":"2025-10-06 16:13:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1065414,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3999552/v1/3d5810f6-70b6-4ee8-b0b9-6163a9322c42.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Automated contouring of gross tumor volume lymph nodes in lung cancer by deep learning","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eIn recent years, lung cancer has become the second most common cancer with the highest mortality rate in the world [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Radiation therapy is an important treatment for lung cancer and the accurate contouring of the gross target volume (GTV) of the lung is essential for the precise delivery of radiotherapy [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Lymph Node (LN) metastasis is common in lung cancer, particularly in inoperable patients. Precision delineation of the GTV lymph nodes (GTVnd) is a necessary step in clinical target delineation for lung cancer radiotherapy and was recommended by the treatment guidelines for lung cancer [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Although positron emission tomography CT (PET-CT) is the most appropriate method by which to identify the GTVnd, not all lung cancer patients undergo PET-CT scanning because of economic and other factors, particularly in underdeveloped regions [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Therefore, it is mainly segmented manually by oncologists in clinical application, highly dependent on manual experience, and results in significant heterogeneity, despite oncologists following the same guidelines [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Moreover, it is particularly difficult for beginners or inexperienced clinical oncologists to delineate the GTVnd. Therefore, it is of great clinical value to achieve automatic and accurate segmentation of the GTVnd in patients with lung cancer with LNs metastasis.\u003c/p\u003e \u003cp\u003eWith the development of deep learning, convolutional neural network (CNN)-based methods have been successfully applied to the automatic contouring of lung cancer. Significant progress has been made in the study of automatic organ-at-risk (OAR) segmentation [\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], the GTV of primary tumors [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], and the clinical target volume (CTV) in postoperative radiation therapy (PORT) [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Ibragimov et al. used CNNs for OARs segmentation in H\u0026amp;N CT images and obtained Dice Similarity Coefficient (DSC) values that varied from 37.4% for the chiasm to 89.5% for the mandible [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]; however, no target was segmented. UNet, proposed by Ronneberger et al. is primarily used for the segmentation of biomedical images [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] and did not specialize in target segmentation in the field of radiotherapy. Recently, Shen et al. proposed the DiUNet model to automatically delineate the CTV of patients with lung cancer with LN metastasis [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Although the mean value of the 3D DSC was greater than 0.7, the GTVnd was manually contoured and added to the CT slice as a second input in the CNN architecture. To the best of our knowledge, no studies have been conducted on the automatic segmentation of the GTVnd in lung cancer.\u003c/p\u003e \u003cp\u003eThe autodelineation of the GTVnd is difficult because of its unclear boundaries and variability in the size and shape. Although UNet, which was based on encoder-decoder architecture, has shown effective performance in conducting multilevel feature fusions, the network is insufficient to represent high-level features, particularly structures that are of significant importance for GTVnd recognition. Therefore, we propose a novel edge-guided contextual cue enhancement network (ECENet) to improve the segmentation performance of the GTVnd by introducing a contextual cue module and incorporating edge guidance into the feature learning of the segmentation network. A contextual cue modeling module is implemented to enhance spatial and semantic information at a distance. To the best of our knowledge, this is the first study to attempt to use this deep learning approach to delineate the GTVnd in lung cancer, hoping to improve the efficiency of clinical work and accuracy of target delineation.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003ePatient Datasets\u003c/h2\u003e\n \u003cp\u003eCT data from 90 patients with small cell lung cancer (SCLC) with involved thoracic regional LN were collected from January 2021 to January 2024. The number of slices per patient ranged from 70 to 105, with a total of 8,758 slices. The data were 512 \u0026times; 512, with a pixel pitch of 1.27 \u0026times; 1.27 mm and a thickness of 5 mm, and they were acquired using a Brilliance CT Big Bore (Siemens Healthcare). The private information of the patients was kept confidential during data collection and processing. GTVnd was defined as the area of the LN stations with a maximum short axis of \u0026gt;\u0026thinsp;1 cm. The GTVnd was delineated in the mediastinal and lung windows according to Radiation Therapy Oncology Group (RTOG) guidelines [\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e]. GTVnd was manually delineated by senior radiation oncologists.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003eArchitecture of ECENet\u003c/h2\u003e\n \u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, based on the encoder-decoder architecture of UNet [\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e], the encoder of the proposed ECENet extracts the multiscale features from the input image, including low-level spatial details and high-level semantic cues. The proposed contextual cue enhancement module is then used to enforce the consistency of the contextual cues encoded in the deepest feature. The proposed ECENet decoder aims to combine the extracted multiscale features. In this process, for the three lowest layers, the mask of the GTVnd and the edge of the GTVnd area are used to supervise and guide the simultaneous learning of segmentation mask prediction and edge prediction. In this manner, using the final learned edge-aware features, an edge-preserving segmentation prediction can be obtained.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e(1) Contextual cues enhancement module\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eThe proposed contextual cue enhancement module consists of convolution kernel prediction and contextual cue aggregation submodules. As shown in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, the convolution kernel prediction submodule is first used to model both the long-distance and short-distance semantic and spatial relations between any two pixels in the manner of matrix multiplication, thereby generating a spatial-adaptive contextual convolution kernel. The generated contextual convolution kernel is then applied to the original feature, leading to aggregated contextual cue enhanced features.\u003c/p\u003e\n \u003cp\u003eConcretely, for the deepest feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{X}\\in {\\text{R}}^{\\text{H}\\times \\text{W}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e, the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{X}\\)\u003c/span\u003e\u003c/span\u003e is first transformed with two independent convolutional operations, generating two features: the key \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{K}\\in {\\text{R}}^{\\text{H}\\times \\text{W}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e and the query \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{Q}\\in {\\text{R}}^{\\text{H}\\times \\text{W}\\times {\\text{s}}^{2}}\\)\u003c/span\u003e\u003c/span\u003e, where H, W and C denote the height, width and number of channels of the features, respectively, and s denotes the scale of the contextual convolution kernel. In order to model the relation between any two pixels, the feature of any pixel locations should be interacted in an explicit manner. To this end, the features of key and query are reshaped into a 2D shape: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({K}^{{\\prime }}\\in {\\text{R}}^{(\\text{H}\\times \\text{W})\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({Q}^{{\\prime }}\\in {\\text{R}}^{（\\text{H}\\times \\text{W}）\\times {\\text{s}}^{2}}\\)\u003c/span\u003e\u003c/span\u003e. For the key feature, the length of the feature vector for every pixel location is C. For the query feature, the length of the feature vector for every pixel location is \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{s}}^{2}\\)\u003c/span\u003e\u003c/span\u003e. To explicitly model the interaction between the feature vector of every pixel location of the key feature and the query feature, the element-wise matrix multiplication between the two feature vectors is calculated:\u003c/p\u003e\n \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e$${\\text{S}}^{{\\prime }}\\left(\\text{i},\\text{j}\\right)=\\sum _{\\text{q}=1}^{\\text{H}\\times \\text{W}}{\\text{Q}}^{{\\prime }}\\left(\\text{q},\\text{i}\\right)\\times \\text{K}{\\prime }(\\text{q},\\text{j})$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{i}=\\text{1,2},\\dots ,{\\text{s}}^{2}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{j}=\\text{1,2},\\dots ,\\text{C}\\)\u003c/span\u003e\u003c/span\u003e. Thus, the contextual cues that enhance feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{S}}^{{\\prime }}\\)\u003c/span\u003e\u003c/span\u003e can be obtained. This operation can be rewritten as follows:\u003c/p\u003e\n \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e$${\\text{S}}^{{\\prime }}={\\text{Q}{\\prime }}^{\\text{T}}\\times \\text{K}{\\prime }$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{S}}^{{\\prime }}\\in {\\text{R}}^{{\\text{s}}^{2}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{Q}{\\prime }}^{\\text{T}}\\)\u003c/span\u003e\u003c/span\u003e is the transpose of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{Q}{\\prime }\\)\u003c/span\u003e\u003c/span\u003e. Then, in order to obtain the final contextual convolution kernel, the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{S}}^{{\\prime }}\\in {\\text{R}}^{{\\text{s}}^{2}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e should be reshaped into \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{S}\\in {\\text{R}}^{\\text{s}\\times \\text{s}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eFurthermore, to generate the spatially adaptive weight factor \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{M}\\)\u003c/span\u003e\u003c/span\u003e, the contextual convolution kernel \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{S}\\)\u003c/span\u003e\u003c/span\u003e and the original input feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{X}\\)\u003c/span\u003e\u003c/span\u003e were calculated using depth-wise convolution. First, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{S}\\in {\\text{R}}^{\\text{s}\\times \\text{s}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e is splits into C kernels, with each kernel being \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{s}\\times \\text{s}\\)\u003c/span\u003e\u003c/span\u003e in shape. Then, these kernels are applied for every feature channel of the input feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{X}\\in {\\text{R}}^{\\text{H}\\times \\text{W}\\times \\text{C}}\\)\u003c/span\u003e\u003c/span\u003e. Finally, the weight factor \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{M}\\)\u003c/span\u003e\u003c/span\u003e is multiplied with the original input feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{X}\\)\u003c/span\u003e\u003c/span\u003e in an element-wise manner to generate the final contextual cue enhanced features.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e(2) Edge-guided feature enhancement\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eThe decoder is designed to effectively combine the multiscale features generated by the encoder in a layer-wise manner. In this process, for the lowest three layers, both the mask of the GTVnd region and the edge of the GTVnd region are used for supervisions. The features of the decoder learn to predict the segmentation mask and edge of the GTVnd simultaneously. In this manner, the learned edge-aware features can generate edge-preserving predictions.\u003c/p\u003e\n \u003cp\u003eTo obtain the ground truth of the edge of the GTVnd, we use the following operations to transform the mask of the GTVnd into the edge of the GTVnd. For each pixel, if the ground truth label of this pixel is different from that of any one of the eight neighboring pixels, then this pixel is an edge pixel.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003eExperiments\u003c/h2\u003e\n \u003cp\u003eComplete CT image data were randomly divided into a training dataset containing 75 cases and a testing dataset containing 15 cases in a ratio of 5:1. ECENet was trained for GTVnd contouring with the 75 cases in the training dataset and it was applied to predict 15 cases in the testing dataset. Stochastic gradient descent was used in the optimization, and the initial learning rate was set to 0.001. An NVIDIA RTX 4090 GPU was used in the experiments. Because ECENet was developed based on UNet, the performance of this model in contouring the GTVnd in lung cancer was comprehensively evaluated by comparing its performance with that of UNet. UNet was trained separately for GTVnd contouring under the same settings as ECENet.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003eQuantitative Evaluation Metrics\u003c/h2\u003e\n \u003cp\u003eTo quantify the contouring accuracy, the Dice Similarity Coefficient (DSC) [\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e] and 95th percentile Hausdorff Distance (95HD) [\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e] were used. The DSC is defined as follows:\u003c/p\u003e\n \u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e$$DSC{\\kern 1pt} {\\kern 1pt} {\\kern 1pt} \\left( {A\\invalidcharacter B} \\right){\\text{=}}\\frac{{2\\left| {A \\cap \\left. B \\right|} \\right.}}{{\\left| A \\right|{\\text{+}}\\left| B \\right|}}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere \u003cem\u003eA\u003c/em\u003e represents the predicted mask, \u003cem\u003eB\u003c/em\u003e the GT mask, and |\u003cem\u003eA\u003c/em\u003e\u0026cap;\u003cem\u003eB\u003c/em\u003e| the intersection of \u003cem\u003eA\u003c/em\u003e and \u003cem\u003eB\u003c/em\u003e. The DSC measured the relative volumetric overlap between two segmented masks, with a higher value indicating a higher overlap ratio and a value of one meaning that the two masks are identical.\u003c/p\u003e\n \u003cp\u003eThe 95 HD is defined as:\u003c/p\u003e\n \u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e$$\\begin{gathered} 95HD\\left( {A,B} \\right){\\text{=percentile}}\\left[ {{\\text{h}}\\left( {A,B} \\right) \\cup {\\text{h}}\\left( {B,A} \\right),95th} \\right] \\hfill \\\\ h\\left( {A,B} \\right)=\\mathop {\\hbox{max} }\\limits_{{a \\in A}} \\mathop {\\hbox{min} }\\limits_{{b \\in B}} \\left\\| {a - b} \\right\\| \\hfill \\\\ h\\left( {B,A} \\right)=\\mathop {\\hbox{max} }\\limits_{{b \\in B}} \\mathop {\\hbox{min} }\\limits_{{a \\in A}} \\left\\| {b - a} \\right\\| \\hfill \\\\ \\end{gathered}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\left\\| . \\right\\|\\)\u003c/span\u003e\u003c/span\u003e represents the Euclidean norm of points \u003cem\u003eA\u003c/em\u003e and \u003cem\u003eB\u003c/em\u003e. \u003cem\u003eA\u003c/em\u003e= {a\u003csub\u003e1\u003c/sub\u003e, a\u003csub\u003e2\u003c/sub\u003e, \u0026hellip;, a\u003csub\u003en1\u003c/sub\u003e} and \u003cem\u003eB\u003c/em\u003e= {b\u003csub\u003e1\u003c/sub\u003e, b\u003csub\u003e2\u003c/sub\u003e, \u0026hellip;, b\u003csub\u003en2\u003c/sub\u003e} represent two finite point sets. The 95HD value indicates the 95th percentile of mismatches between \u003cem\u003eA\u003c/em\u003e and \u003cem\u003eB\u003c/em\u003e. The 95HD reflected the alignment between the wo contours, with a higher value indicating a larger difference. The DSC and 95HD were calculated at the 3D level, and the mean standard deviation (SD) values were calculated by averaging the values.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"RESULTS","content":"\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shown the GTVnd segmentation slice of one patient, including the transversal, coronal, and sagittal sections. The GTVnd was located in zones 2 and 4 of the lung cancer LN station, with a volume of 137.22 cc. The 3D DSC of this case from UNet and ECENet were 0.40 and 0.83, respectively, and the 95HD values were 6.93 mm, and 8.54 mm, respectively. Compared to the UNet models, ECENet performed much better in the segmentation of the GTVnd, which was highly consistent with the oncologist's manual segmentation contours.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe 3D DSC and 95HD were calculated at a 3D level for the testing dataset and the mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (SD) values for UNet and ECENet were calculated by averaging the values. The quantitative evaluation results were summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The mean 3D DSCs for UNet and ECENet were 0.46\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17 and 0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09, respectively, and the mean 95HDs were 12.76\u0026thinsp;\u0026plusmn;\u0026thinsp;13.12 mm and 12.65\u0026thinsp;\u0026plusmn;\u0026thinsp;5.82 mm, respectively. The proposed ECENet tended to achieve higher accuracy than did UNet by achieving greater 3D DSC values and lower 95HD values. Moreover, the proposed model performed more robustly on the testing dataset, displaying a small average standard deviation. There was a statistically significant difference between the different models for both 3D DSC (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and 95HD (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001). In addition, the mean volume of GTVnd was 46.06\u0026thinsp;\u0026plusmn;\u0026thinsp;42.33 cc. The standard deviation values are large, owing to the large differences in the GTVnd volume for each case in the testing dataset.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe statistics of 3D DSC and 95HD with the volume of the GTVnd for UNet and ECENet in the testing dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eDSC\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95HD\u0026thinsp;\u0026plusmn;\u0026thinsp;SD (mm)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e0.46\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e12.76\u0026thinsp;\u0026plusmn;\u0026thinsp;13.12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eECENet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e12.65\u0026thinsp;\u0026plusmn;\u0026thinsp;5.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eFew studies have been conducted on the automatic delineation of the GTVnd for LN metastasis in lung cancer, and oncologists mainly perform manual contours based on their clinical experience, which leads to heterogeneity between different oncologists and intraobservers at different times [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. However, consistent target delineation is essential for improving radiotherapy outcomes. Since the artificial intelligence successfully performed automatic target delineation, particularly in improving the accuracy of contours and reducing heterogeneities [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], an ECENet based on UNet was constructed to automatically delineate the contours of the GTVnd in lung cancer accurately and consistently. The segmentation performance of the model was also evaluated.\u003c/p\u003e \u003cp\u003eThe DSC value was commonly used to assess the accuracy of autosegmentation during radiation treatment planning. Because ECENet was developed based on UNet, its performance in contouring the GTVnd in lung cancer was comprehensively evaluated by comparing it with UNet. The experimental results in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e demonstrated that the proposed outperforms UNet models. The DSC values increased from 0.46 to 0.72, revealing that it is feasible and effective to develop ECENet for segmenting GTVnd contours in lung cancer by implementing a contextual cue enhancement module and edge-guided feature enhancement. Moreover, for the segmentation of the GTVnd, the DSC values have been reported to be 0.46 [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] and 0.62 [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], whereas our proposed model demonstrated a high DSC value of 0.72. These models were not tested using the same dataset; therefore, it would be unfair to say that our proposed model was superior. However, it was reasonable to conclude that ECENet yields good results.\u003c/p\u003e \u003cp\u003eAlthough the segmentation accuracy of GTVnd is better than the previously reported, it still displays inferior performance compared with the published GTV or OAR contouring models [\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], in which the mean DSC values tend to be greater than 0.8. This reason was mainly because the GTV and OAR were characterized differently from the GTVnd. While most GTV or OAR regions had well-defined boundaries, the GTVnd did not, and the contour boundaries of GTVnd could easily be confused with the various complex anatomical structures around it. Moreover, there were considerable differences in the shape, volume, and location of the GTVnd among patients. As a result, the more complex GTVnd contouring performed inferiorly compared to the GTV or OAR contouring models.\u003c/p\u003e \u003cp\u003eThis was a single-center study with a small sample size, and the segmentation accuracy of ECENet for GTVnd was inferior compared to that of the reported GTV or OAR contouring. However, this model is superior to UNet and shows a higher DSC value than the reported model for the segmentation of the GTVnd. Since Zijdenbos et al. suggested that a DSC value of \u0026gt;\u0026thinsp;0.7 represents good overlap [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], this model with a DSC value of 0.72 can accurately achieve the automatic contouring of GTVnd in lung cancer. However, the segmentation accuracy of the GTVnd remains unsatisfactory for clinical application. Moreover, the dose deviation due to imperfect GTVnd segmentation needs to be further calculated, and its impact on treatment efficacy should be evaluated to avoid the under dosage of target volumes or an overdose being delivered to normal tissues.\u003c/p\u003e \u003cp\u003eIn general, this model can provide a potential benefit for the automatic contouring of GTVnd and CTV in lung cancer. In the future, the dataset size could be expanded to further optimize the model and improve its applicability and overall accuracy. With further revision of the dose deviation due to imperfect GTVnd segmentation in treatment planning, the model will be abled able to be clinically applied. Simultaneously, this technology can be explored for the automatic segmentation of the GTVnd in other types of tumors.\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eThe automatic and accurate segmentation of the GTVnd is key to achieving the automatic delineation of the CTV for lung cancer with LN metastasis. In this study ECENet was proposed for the fully automatic and accurate contouring of GTVnd in lung cancer. The model was based on UNet with edge guidance and spatial context enhancement. The results of the quantitative evaluation this model showed that it could accurately delineate the contours of the GTVnd. However, a multicenter clinical evaluation with more cases is needed to fully apply the model in clinical practice. In conclusion, this study explored the application of deep learning to the automatic delineation of the GTVnd in lung cancer. ECENet has shown the potential to improve the consistency of contouring, which also provides the potential to positively contribute to the automatic delineation of the GTVnd.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by Jiangxi Cancer Hospital Doctoral Research Start-up Fund (No. BSQDJ202301) and\u0026nbsp;the Jiangxi Cancer Hospital Scientific Research Open Fund Project (No. KFJJ2023YB19).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analyzed during this study are included in this published article and its supplementary information files.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the ethics committee of the Jiangxi Cancer Hospital (2023ky192). Informed consent was obtained from all patients included in this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eSung H, Ferlay J, Siegel R L, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA: A Cancer Journal for Clinicians, 2021, 71(3): 209-249.\u003c/li\u003e\n \u003cli\u003eVan Rikxoort E M, de Hoop B, Viergever M A, et al. Automatic lung segmentation from thoracic computed tomography scans using a hybrid approach with error detection[J]. Medical physics, 2009, 36(7):2934\u0026ndash;2947.\u003c/li\u003e\n \u003cli\u003eEttinger DS, Wood DE, Akerley W, Bazhenova LA, Borghaei H, Camidge DR, et al. Non-small cell lung cancer, version 6.2015. J National Comprehensive Cancer Network.2015;13(5):515\u0026ndash;24.\u003c/li\u003e\n \u003cli\u003eDe Ruysscher D, Faivre-Finn C, Nestle U, Hurkmans CW, Le P\u0026eacute;choux C, Price A, et al. European organisation for research and treatment of cancer recommendations for planning and delivery of high-dose, high-precision radiotherapy for lung cancer. J Clin Oncol. 2010;28(36):5301\u0026ndash;10.\u003c/li\u003e\n \u003cli\u003ePetersen H, Holdgaard PC, Madsen PH et al (2016) FDG PET/CT in cancer: comparison of actual use with literature-based recommendations. Eur J Nucl Med Mol Imaging 43:695\u0026ndash;706.\u003c/li\u003e\n \u003cli\u003eFletcher JW, Djulbegovic B, Soares HP et al (2008) Recommendations on the use of 18F-FDG PET in oncology. J Nucl Med 49:480\u0026ndash;508.\u003c/li\u003e\n \u003cli\u003eBenjamin E. Nelms, Wolfgang A. Tom\u0026eacute;, Greg Robinson, James Wheeler. Variations in the Contouring of Organs at Risk: Test Case From a Patient With Oropharyngeal Cancer[J]. International Journal of Radiation Oncology, Biology, Physics, 2010, 82(1):368-378.\u003c/li\u003e\n \u003cli\u003eCoen Rasch, Avraham Eisbruch, et al. Irradiation of paranasal sinus tumors, a delineation and dose comparison study[J]. International Journal of Radiation Oncology, Biology, Physics, 2002, 52(1):120-7.\u003c/li\u003e\n \u003cli\u003eUzelaltinbulat S, Ugur B. Lung tumor segmentation algorithm[J]. Procedia Computer Science, 2017, 120: 140-147.\u003c/li\u003e\n \u003cli\u003eSoltani-nabipour J, Khorshidi A, Noorian B. Lung tumor segmentation using improved region growing algorithm[J]. Nuclear Engineering and Technology, 2020, 52(10): 2313-2319.\u003c/li\u003e\n \u003cli\u003eYang B L, Xiang D H, Yu F H, et al. Lung tumor segmentation based on the multi-scale template matching and region growing[C]//Proc. SPIE 10578, Medical Imaging 2018: Biomedical Applications in Molecular, Structural, and Functional Imaging, 2018: 105782Q.\u003c/li\u003e\n \u003cli\u003eBallangan C, Wang X, Fulham M, et al. Lung tumor segmentation in PET images using graph cuts[J]. Computer Methods and Programs in Biomedicine, 2013, 109(3): 260-268.\u003c/li\u003e\n \u003cli\u003eMi H M, Petitjean C, Dubray B, et al. Automatic lung tumor segmentation on PET images based on random walks and tumor growth model[C]//IEEE International Symposium on Biomedical Imaging, 2014: 1385-1388.\u003c/li\u003e\n \u003cli\u003eNian Y J, Li M, Cui H R, et al. Graph-based unsupervised segmentation for lung tumor CT images[C]//IEEE International Conference on Computer and Communications, 2017: 1884- 1888.\u003c/li\u003e\n \u003cli\u003eOda H, Bhatia K K, Roth H R, et al. Dense Volumetric Detection and Segmentation of Mediastinal Lymph Nodes in Chest CT Images[C]. MORI K, PETRICK N, eds. Medical Imaging 2018: Computer Aided Diagnosis. SPIE, 2018: 1.\u003c/li\u003e\n \u003cli\u003eIbragimov B, Xing L. Segmentation of organs-at-risks in head and neck CT images using convolutional neural networks. Med Phys (2017) 44(2):547\u0026ndash;57. doi:10.1002/mp.12045.\u003c/li\u003e\n \u003cli\u003eRonneberger O, Fischer P, Brox T. UNet: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical Image Computing and Computer Assisted Intervention, 2015, 9351: 234-241.\u003c/li\u003e\n \u003cli\u003eJie Shen, Fuquan Zhang, Mingyi Di, et al. Clinical target volume automatic segmentation based on lymph node stations for lung cancer with bulky lump lymph nodes. Thorac Cancer. 2022 Oct; 13(20): 2897\u0026ndash;2903.\u003c/li\u003e\n \u003cli\u003eBradley JD, Paulus R, Komaki R, et al. Standard-dose versus high-dose conformal radiotherapy with concurrent and consolidation carboplatin plus paclitaxel with or without cetuximab for patients with stage IIIA or IIIB non-small-cell lung cancer (RTOG 0617): a randomised, two-by-two factorial phase 3 study. Lancet Oncol. 2015;16(2):187\u0026ndash;99.\u003c/li\u003e\n \u003cli\u003eRonneberger O, Fischer P, Brox T. UNet: Convolutional networks for biomedical image segmentation. arXiv preprint arXiv:1505.04597, 2015\u003c/li\u003e\n \u003cli\u003eDice LR. Measures of the amount of ecologic association between species. Ecology. 1945;26(3):297\u0026ndash;302.\u003c/li\u003e\n \u003cli\u003eHuttenlocher DP, Klanderman GA, Rucklidge WJ. Comparing images using the Hausdorff distance. IEEE Trans Pattern Anal Mach Intell. 1993;15(9):850\u0026ndash;63.\u003c/li\u003e\n \u003cli\u003eMercieca S, Belderbos JSA, van Baardwijk A, et al. The impact of training and professional collaboration on the interobserver variation of lung cancer delineations: a multi-institutional study. Acta Oncol. 2019;58(2):200\u0026ndash;8.\u003c/li\u003e\n \u003cli\u003eGuangying Z, Tingyi X, Lvhua W, et al. Consensus and controversies on delineation of radiotherapy target volume for patients with non-small cell lung cancer. Chin J Rasiat Oncol. 2008;17(6):432\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eBarkati M, Simard D, Taussky D, et al. Magnetic resonance imaging for prostate bed radiotherapy planning: An inter- and intra-observer variability study. J Med Imaging Radiat Oncol (2016) 60:255\u0026ndash;9. doi: 10.1111/1754-9485.12416.\u003c/li\u003e\n \u003cli\u003eYang J, Beadle BM, Garden AS, et al. Auto-segmentation of low-risk clinical target volume for head and neck radiation therapy. Pract Radiat Oncol (2014) 4(1): e31\u0026ndash;7. doi:10.1016/j.prro.2013.03.003\u003c/li\u003e\n \u003cli\u003eKuo Men, Xinyuan Chen, Ye Zhang, et al. Deep Deconvolutional Neural Network for Target Segmentation of Nasopharyngeal Cancer in Planning Computed Tomography Images. Front Oncol. 2017 Dec 20;7:315. doi: 10.3389/fonc.2017.00315. PMID: 29376025; PMCID: PMC5770734.\u003c/li\u003e\n \u003cli\u003eThibault Marin, Yue Zhuo, Rita Maria Lahoud, et al. Deep learning-based GTV contouring modeling inter- and intra- observer variability in sarcomas, Radiotherapy and Oncology, Volume 167, 2022, Pages 269-276, ISSN 0167-8140, https://doi.org/10.1016/j.radonc.2021.09.034.\u003c/li\u003e\n \u003cli\u003eLang, D.M., Peeken, J.C., Combs, S.E.,et al. Deep Learning Based GTV Delineation and Progression Free Survival Risk Score Prediction for Head and Neck Cancer Patients. Lecture Notes in Computer Science, vol 13209. Springer, Cham. https://doi.org/10.1007/978-3-030-98253-9_14.\u003c/li\u003e\n \u003cli\u003eShi, F., Hu, W., Wu, J. et al. Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. Nat Commun 13, 6566 (2022). https://doi.org/10.1038/s41467-022-34257-x.\u003c/li\u003e\n \u003cli\u003eZijdenbos AP, Dawant BM, Margolin RA, Palmer AC. Morphometric analysis of white matter lesions in MR images: method and validation. IEEE Trans Med Imaging (1994) 13(4):716\u0026ndash;24. doi:10.1109/42.363096\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Automatic segmentation, Gross target volume lymph nodes (GTVnd), Lung cancer, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-3999552/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3999552/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eThe precise contouring of gross tumor volume lymph nodes (GTVnd) is an essential step in clinical target volume delineation. However, to the best of our knowledge, there have been no autosegmentation studies on the GTVnd in lung cancer. This study aims to propose and evaluate a deep learning model for segmenting GTVnd in lung cancer.\u003c/p\u003e\u003ch2\u003eMethod\u003c/h2\u003e \u003cp\u003eNinety computed tomography (CT) scans of patients with lymph node metastasis in lung cancer were collected, of which 75 patients were assembled into a training dataset and 15 were used in a testing dataset. A new segmentation model was constructed to enable the automatic and accurate delineation of the GTVnd in lung cancer. This model integrates a contextual cue enhancement module and an edge-guided feature enhancement decoder. The contextual cues enhancement module was used to enforce the consistency of the contextual cues encoded in the deepest feature, and the edge-guided feature enhancement decoder was used to obtain edge-aware and edge-preserving segmentation predictions. The model was quantitatively evaluated using the three-dimensional Dice Similarity Coefficient (3D DSC) and the 95th Hausdorff Distance (95HD).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe mean 3D DSC value of the ECENet was approximately 0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09, and that of the 95HD was approximately 12.65\u0026thinsp;\u0026plusmn;\u0026thinsp;5.82 mm. The performance of ECENet was significantly improved, compared with UNet (0.46\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19 and 12.76\u0026thinsp;\u0026plusmn;\u0026thinsp;13.12 mm, respectively). There were statistically significant differences in terms of the 3D DSC and 95HD values between from ECENet and UNet.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe proposed model could achieve the automatic delineation of the GTVnd in the thoracic region of lung cancer and showed certain advantages, making it a potential choice for the automatic delineation of the GTVnd in lung cancer, particularly for young radiation oncologists.\u003c/p\u003e","manuscriptTitle":"Automated contouring of gross tumor volume lymph nodes in lung cancer by deep learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-03-19 16:14:35","doi":"10.21203/rs.3.rs-3999552/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-03-15T06:51:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-03-15T06:33:41+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-03-15T06:33:41+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Cancer","date":"2024-02-29T11:46:42+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7eaa54b9-367f-46b6-9121-239b50d7a6b5","owner":[],"postedDate":"March 19th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-10-06T16:12:28+00:00","versionOfRecord":{"articleIdentity":"rs-3999552","link":"https://doi.org/10.1186/s12885-025-14794-6","journal":{"identity":"bmc-cancer","isVorOnly":false,"title":"BMC Cancer"},"publishedOn":"2025-09-30 15:58:17","publishedOnDateReadable":"September 30th, 2025"},"versionCreatedAt":"2024-03-19 16:14:35","video":"","vorDoi":"10.1186/s12885-025-14794-6","vorDoiUrl":"https://doi.org/10.1186/s12885-025-14794-6","workflowStages":[]},"version":"v1","identity":"rs-3999552","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3999552","identity":"rs-3999552","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}