Precision in radiotherapy. Radiation therapy for lung cancer requires accurate mapping of the gross tumor volume lymph nodes (GTVnd) - the regions where cancer has spread to nearby lymph nodes. These contours directly determine where high-dose radiation is delivered, making accuracy essential for both tumor control and sparing healthy tissue.
The variability problem. Currently, GTVnd contouring is done manually by radiation oncologists, which creates substantial inconsistency. Even when following the same clinical guidelines, different physicians produce significantly different contours - a problem known as inter-observer variability. This variability is especially pronounced among less experienced physicians.
Limited access to advanced imaging. PET-CT scanning provides the most accurate lymph node identification, but economic barriers and regional resource limitations mean many lung cancer patients - particularly in developing regions - receive only standard CT imaging. This constrains the quality of contouring available to many patients globally.
An underexplored problem. While deep learning has been applied to segment primary tumors and organs at risk in radiotherapy, automated segmentation of GTVnd specifically for lung cancer remains a relatively unexplored area. This study represents one of the early systematic attempts to address this gap.
Building on UNet. The researchers developed ECENet (Edge-guided Contextual cue Enhancement Network) starting from the established UNet encoder-decoder architecture, which has proven effective for medical image segmentation. ECENet adds two key innovations on top of this foundation to better handle the unique challenges of lymph node delineation.
Contextual cue enhancement module. The first major addition captures relationships between distant parts of the image. Lymph nodes often lack clear boundaries and can be confused with surrounding structures, so understanding the broader spatial context of the image helps the model make better predictions. This module uses a self-attention-inspired approach to model long-distance and short-distance pixel relationships simultaneously.
Edge-guided feature enhancement decoder. The second innovation trains the network to simultaneously learn both the segmentation mask (the full region of the GTVnd) and the boundary edges of that region. By supervising edge prediction alongside mask prediction, the decoder learns edge-aware features that produce sharper, more accurate contour boundaries - a critical requirement for clinical use.
Dual loss functions. The model is trained with two simultaneous objectives: a standard binary segmentation loss for the overall mask, and an edge loss specifically penalizing errors at the GTVnd boundaries. This dual supervision directly addresses the core difficulty of the task - accurately capturing unclear and variable lymph node margins.
Study cohort. The study enrolled 90 patients with histologically confirmed stage III-IV small cell lung cancer (SCLC) treated with concurrent chemoradiotherapy between January 2021 and January 2024 at Jiangxi Cancer Hospital in China. All patients received intensity-modulated radiotherapy (IMRT) at 60 Gy in 30 fractions combined with platinum-based chemotherapy.
Data characteristics. CT scans were acquired on a Siemens Brilliance CT Big Bore scanner, generating images of 512 x 512 pixels with a pixel pitch of 1.27 x 1.27 mm and 5 mm slice thickness. Each patient had 70 to 105 CT slices, totaling 8,758 slices across all patients. GTVnd was defined as lymph node regions with a maximum short axis greater than 1 cm.
Ground truth annotation. GTVnd contours were manually delineated by senior radiation oncologists following RTOG guidelines, then reviewed by two additional oncologists each with more than 10 years of specialized expertise in pulmonary radiotherapy. This rigorous expert annotation process established high-quality ground truth for training and evaluation.
Training and testing split. Data were divided with 75 patients for training and 15 for testing. ECENet was compared against two established baselines: standard UNet and nnUNet, a state-of-the-art self-configuring framework. All models were trained under identical experimental conditions for fair comparison.
ECENet outperforms competitors. On the test set, ECENet achieved a mean 3D Dice Similarity Coefficient (DSC) of 0.72 +/- 0.09 and a 95th percentile Hausdorff Distance (95HD) of 6.39 +/- 4.59 mm. In comparison, standard UNet scored a DSC of 0.46 +/- 0.19 with a 95HD of 12.24 +/- 13.36 mm, and nnUNet achieved 0.52 +/- 0.18 with a 95HD of 9.92 +/- 6.49 mm.
Meaning of the metrics. The DSC measures how much the predicted contour overlaps with the expert-drawn ground truth - a score of 1.0 means perfect agreement. The 95HD measures the distance between contour boundaries, where lower values indicate better shape accuracy. ECENet's improvements on both metrics were statistically significant (P less than 0.001 for both).
Compared to physician performance. ECENet's accuracy placed it between junior physicians (DSC 0.68 +/- 0.10) and mid-level physicians (DSC 0.81 +/- 0.06). Senior physicians achieved the highest accuracy at 0.90 +/- 0.02. This positions ECENet as a meaningful tool for supporting less experienced clinicians while still falling short of expert-level performance.
Robustness improvement. ECENet also demonstrated substantially lower variability across test cases compared to competing models, as reflected in its smaller standard deviations. This consistency is important for clinical applications where reliable performance across diverse patient presentations is essential.
Beyond geometric accuracy. Geometric metrics like DSC measure contour overlap, but the ultimate test of a segmentation tool is whether it produces treatment plans that are clinically equivalent to expert-drawn plans. The researchers compared radiotherapy plans generated from ECENet contours against plans based on manual clinical contours.
Excellent PTV dose agreement. Dosimetric analysis of the planning target volume (PTV) revealed very close agreement, with average relative differences of less than 0.17% for the key dose metrics D2, D50, and D98. The homogeneity and conformity indices were also statistically equivalent between auto-contoured and manually contoured plans.
Normal tissue dose comparability. Radiation doses to critical nearby organs were also closely matched. For the lungs, dose differences were within 3.5% across all measured parameters. Heart dose differences were within 6.1% for most parameters, with V20 for heart showing slightly larger differences at 10.5%. All differences were non-significant (P greater than 0.05).
TCP and NTCP concordance. The tumor control probability (TCP) - a model-based estimate of treatment effectiveness - showed near-identical results between clinical (66.99%) and predicted (66.88%) plans. The normal tissue complication probability (NTCP) for lung was similarly close at 3.13% versus 3.25%. These biological endpoint agreements confirm the clinical relevance of ECENet-generated contours.
Why GTVnd is harder than other targets. GTVnd segmentation consistently yields lower DSC values than primary tumor (GTV) or organ-at-risk (OAR) segmentation, which typically achieve DSC values above 0.8. The researchers attribute this to the unclear boundaries of lymph nodes, their high variability in shape and size, and their proximity to complex anatomical structures that can appear visually similar on CT.
ECENet's position in prior literature. Previous reported models for GTVnd segmentation achieved DSC values of approximately 0.46 and 0.62. ECENet's 0.72 places it above these benchmarks, and a DSC above 0.7 is generally considered to represent good overlap in the medical imaging literature. This suggests clinically meaningful improvement, though direct comparison is complicated by differences in datasets.
The dosimetric explanation for plan consistency. The authors note that despite ECENet contours differing from manual contours in some cases, the resulting treatment plans were clinically equivalent. This appears to be because GTVnd comprises only a fraction of the total planning target volume - the primary tumor CTV dominates the plan, which dilutes the impact of lymph node contour variations on overall dose distribution.
Limitations acknowledged. This was a single-center study with 90 patients, limiting generalizability. ECENet does not simultaneously segment the primary tumor, which would be needed for a complete automated radiotherapy planning workflow. The study also did not assess intra-observer variability. Multicenter validation with larger datasets is needed before clinical adoption.
A tool for junior physicians. ECENet performs at a level between junior and mid-level radiation oncologists in accuracy, making it particularly valuable as a first-pass tool for less experienced physicians. It can generate initial contours that junior physicians then review and refine, reducing workload while providing a quality baseline and reducing variability.
Reducing inter-observer variability. One of the most clinically significant benefits of automated contouring is its consistency - the same algorithm produces the same result regardless of time of day, physician fatigue, or institutional practice differences. This standardization can improve radiotherapy quality across centers, particularly in resource-limited settings that lack access to senior specialists.
Pathway to full automation. The authors envision future development of a comprehensive framework that jointly segments both primary tumors and GTVnd from the same CT scan. Such a system would provide the complete contour set needed for automated treatment planning, bringing clinical AI closer to end-to-end radiotherapy workflow automation.
Global equity implications. Many regions that cannot access PET-CT imaging for superior lymph node detection still perform CT-based radiotherapy planning. An automated, accurate CT-based GTVnd segmentation tool like ECENet could improve the quality of radiotherapy planning in these settings without requiring expensive additional imaging infrastructure.
Proven feasibility. This study demonstrates that deep learning can achieve clinically acceptable accuracy for the automated contouring of GTVnd in lung cancer using standard CT images alone. ECENet achieved a DSC of 0.72, which crosses the threshold for good geometric overlap, and produced treatment plans dosimetrically equivalent to those drawn by clinical experts.
Architectural innovations validated. The two key innovations of ECENet - contextual cue enhancement and edge-guided learning - successfully addressed the specific challenges of GTVnd segmentation. The model's superior performance over UNet and nnUNet confirms that these targeted additions provide meaningful benefit for this specific task.
Next steps. The research team plans to expand the dataset by incorporating more patients from multiple centers, which will improve model robustness and allow broader validation. Extending the approach to GTVnd segmentation in other cancer types is also proposed, given the same challenges of unclear boundaries and variable lymph node presentation exist across oncology.
A stepping stone for automation. The study's dosimetric validation confirms that automated GTVnd contouring can generate treatment-ready plans with minimal additional intervention. As such, ECENet represents an important building block toward the goal of fully automated, AI-assisted radiation treatment planning for lung cancer patients.