Currently, there is no accurate definition of LN station boundaries, and oncologists mainly delineate CTV in lung cancer based on their clinical experience, resulting in great heterogeneity. When the CTV region is described, neither the specific LN regions contained within nor their range boundaries can be accurately determined. Many studies have found this problem, but none have found an effective solution. Zhu Guangying and his colleague investigate the consensus and controversies on delineation of radiotherapy target volume for patients with NSCLC in 10 radiation departments in china and 2 departments in US. The delineation of the CTV in mediastinal lymph nodes varied greatly 23. In Femke O B Spoelstra’ s study in 2010, seventeen thoracic radiation oncologists were invited to contour their routine CTV for 2 representative NSCLC patients, the delineation of the CTV was vary greatly between different doctors at different times. As showed in Supplementary Figure S3, CTV were contoured at an interval of 2 weeks by six oncologists 21. Another study in 2019 also confirmed this view, image interpretational differences can lead to large interobserver variation particularly when delineating the gross tumor volume lymph node 22. In other words, even though the CTV containes same lymph node zones, the boundaries of each region vary greatly in different oncologist. The heterogeneity in the CTV is particularly evident when there is bulky lump GTVnd in the LN stations. Although automatic segmentation models for CTV in lung cancer have been proposed, they failed in cases involving bulky lump GTVnd. In this study, the boundaries of each LN region were accurately defined for the first time in RT, and a CNN-based model was trained to auto contour LN stations for accurate and consistent CTV formation, which is an important innovation point of this study.
The boundaries of LN stations vary widely in shape, location, and size of GTVnd changes. Therefore, the prior GTVnd expert knowledge was incorporated as an additional training condition of the proposed DiUnet model, which constrained the output mask boundaries using the spatial features extracted from the GTVnd boundaries, leading to more accurate results.
Examples of the LN stations contoured using DiUNet are shown in Fig. 3, in which LN boundaries were contoured accurately even for bulky lump GTVnd. This has great value in CTV contouring, especially for central lung carcinoma patients. Therefore, it was first applied to auto-contouring with machine learning models for lung cancer.
The DSC values of DiUNet in each LN station were greater than 0.7, except for Station 8 (Fig. 4). The inaccuracy of contouring the oesophagus may explain this lower value.
The percentage of clinically acceptable scores was greater than 98% and the average score was higher than 2.91 in the DiUNet model (Table 2). LN stations contoured by AI were acceptable compared with human-generated structures. The CTV evaluation results are presented in Table 3. Despite the differences in the evaluation systems among physicians, it is believed that the CTV contoured by DiUNet can be clinically acceptable. DiUNet passed the consistency test with an overall positivity rate higher than 50% in the comparison of LN stations and CTV. The concordance of data between the two oncologists was poor, but they both considered that LN stations and the CTV contoured with AI models were not inferior to those of MC (Tables S4 and S5).
Accurate definition and complete delineation of lymph node stations’ boundaries are an important basis for the formation of standardized CTV. This is the first study to delineate lymph node stations to promote CTV standardization.
It should be noted that most slices that required corrections were located at the edges adjacent to the blood vessel or at the border of the LN station. This may be because accurate delineation needs to integrate information from multiple upper and lower slices, but the information available at the boundary is limited. In the future, we are planning to explore the impact of CTV on survival prognoses that are comprised of intact partially involved LN regions, local recurrence, and rate of nodal failures.