ESTRO 2024 - Abstract Book

S480

Clinical - Breast

ESTRO 2024

A cohort of 66 left-sided breast cancer patients treated with radiotherapy from 2017-22 from a single institution in the Netherlands was included. Data collection was approved by the institutional review board. CT scans and the clinical structure sets were collected, where target delineation was done according to ESTRO guidelines (IMN_CLIN). The delineations were done by the treating radiation oncologist and represent the interpretation of the ESTRO guidelines at that institute. A second set of IMN delineations (IMN_DL) was created with a deep learning model (3D full resolution nnUNet [2]) that was trained on multi-institutional consensus-based delineations (based on the ESTRO consensus guideline) from the Danish Breast Cancer Group (DBCG). The model performed on par with experts [3] within the DBCG group and as such represents the Danish interpretation of the ESTRO consensus guidelines. The geometrical similarity between the IMN_CLIN and IMN_DL was evaluated using the Dice Similarity Coefficient (DSC), the Hausdorff 95th percentile distance (HD95) and the mean surface distance (MSD). The lateral width of the IMN, the smallest distance to sternum and the maximum distance between the IMNs in the medial direction in slices with the heart structure present were measured and the results were displayed in medians with interquartile ranges (IQR). Treatment plans were generated for both IMN_DL and IMN_CLIN using Pinnacle Auto-Planning and scripting. A clinical template plan was used with slight modifications to meet the requirements of this study. All plans were rescaled to V95%=99% target coverage. The plans used 6MV dual arc VMAT beams, 40.05 Gy in 15 fractions. Dose sparing objectives were set for heart, lungs, contralateral breast, and rings for conformity. All IMNs were cropped to the same caudal border (the most cranial of the two caudal border), to avoid differences arising from caudal nodal inclusion criteria. Differences in MHD were compared and correlated with the geometrical similarity metrics described above to highlight the influence of guideline interpretation. Statistical significance was tested with the Wilcoxon signed rank test with significance level 0.05, and the Pearson correlation test. Four patients were removed due to various data errors, leaving a total of 62 patients for analysis. The geometrical similarity between the IMN_DL and IMN_CLIN gave a median DSC = 0.63 (IQR 0.57-0.68), median HD95% = 4.89 mm (IQR 3.15 mm-6.83 mm) and a median MSD = 0.85 mm (IQR 0.69 mm-1.13 mm). The IMN_CLIN were wider (median width = 18.2 mm) and delineated closer to sternum (median distance = 4.3 mm) compared to the IMN_DL (median width = 14.7 mm and median distance = 7.3 mm) all p << 0.05, see figure 1. The MHD was significantly different (p=0.01) between the plans created with the IMN_DL and the IMN_CLIN, with a median MHD = 5.07Gy (IQR 4.14 Gy-5.76Gy) for IMN_CLIN and a median MHD = 4.68 Gy (IQR 4.03Gy-5.67Gy) for IMN_DL. In 11/62 patients, the dose difference was larger than 0.5 Gy and the largest dose difference was 0.9 Gy. For 38/62 cases, the IMN_CLIN showed a higher MHD, see figure 2. The strongest correlation was found between the MHD difference and the max distance in the medial direction, with Pearson coefficient = -0.26 (p = 0.06), see figure 2. No correlation between IMN volume and MHD (Pearson coefficient = -0.06, p= 0.61) and between the distance to sternum and MHD (Pearson coefficient = -0.09, p = 0.46) were found. Results: