S1571
Physics - Autosegmentation
ESTRO 2026
Digital Poster 2399
Qualitative and quantitative evaluation of auto- segmentated clinical target volumes in breast cancer Love Dahlstedt Hassler 1 , Pehr Lind 1,2 , Camilla Wendt 1 , Rafat Kojoj 1 , Albert Siegbahn 1,2 1 Oncology, Södersjukhuset, Stockholm, Sweden. 2 Department of Clinical Science and Education Södersjukhuset, Karolinska Institutet, Stockholm, Sweden Purpose/Objective: In postoperative radiotherapy (RT) for breast cancer, manual contouring (MC) of clinical target volumes (CTVs) is resource consuming. Use of deep-learning based auto-segmentation tools could reduce the need of MC. Our study aimed to evaluate the generated structures of two auto-segmentation tools with quantitative data analysis and qualitative assessment of breast cancer CTVs and quantitative metrics of organs of interest (OOIs). Material/Methods: Image data and available manual contours from treatment planning were collected from ten clinical cases of postoperative RT to the breast or chest wall. A physician delineated MC if unavailable, with any new MC-CTV peer-reviewed prior to analysis. Auto- segmentations produced by the auto-contouring software tools AI-Rad Companion Organs RT (Siemens) and Contour+ (MVision) were assessed. The included CTV structures were the breast/chest wall, axillary lymph nodes (LN), and intramammary lymph nodes (IMNs), based on ESTRO guidelines1. Additionally, OOIs relevant to breast cancer RT were included in the quantitative analysis. The used quantitative metrics for CTVs and OOIs, with MC as reference, were volumetric/surface dice similarity coefficient [vDSC/sDSC (2mm)] and maximum/average Hausdorff distance (HD/aHD). Three experienced radiation oncologists performed qualitative evaluations of both MC and auto-segmented CTVs using a five-point Likert scale. Furthermore, we analysed association between quantitative metrics and qualitative results using Kendall’s test. Results: Quantitative metrics are presented in Table 1. Contour+ outperformed AI-Rad in both breast and LN segmentations across all metrics. The metrics for IMN and compared OOIs were largely similar between the two AI tools.
For the evaluated CTV structures, the median Likert scores assigned by reviewers are presented in Table 2. The Contour+ volumes of the breast and LN were mostly accepted or required minor revisions, scoring slightly lower than the MC volumes, although this difference was not statistically significant. The AI-Rad Breast and LN structures scored significantly lower than the MC and often required major revisions. For the IMN, scores were acceptable, with no significant difference between the AI-IMN and MC-IMN volumes.
We also found significant correlations between Likert scores and the quantitative metrics, in breast for sDSC, HD and aHD results and for LN across all metrics.
Made with FlippingBook - Share PDF online