ESTRO 2025 - Abstract Book

S2426

Physics - Autosegmentation

ESTRO 2025

Imaging Archive.[3]Kong L, Sun J, Zhang C. Sde-net: Equipping deep neural networks with uncertainty estimates[J]. arXiv preprint arXiv:2008.10546, 2020.

852

Digital Poster Evaluation of an AI Algorithm for Breast Cancer Radiotherapy Target Delineation within a Multi-institutional Network Rahul Lal lal chowdhary 1 , Manikumar Singamsetty 1 , Suresh Chaudhari 1 , Vibhor Gupta 1 , Harjot Bajwa 1 , sushil beriwal 2,3 1 Radiation oncolohy, American oncology institute, hyderabad, India. 2 Radiation oncolohy, AHN, Pittsburgh, USA. 3 Medical affairs, Varian medical system, wexford, USA Purpose/Objective: Breast cancer regional node radiotherapy poses challenges in delineating target volumes due to oncologists' diverse interpretations of guidelines, resulting in varying contouring practices and substantial time requirements. AI algorithms present an opportunity to standardize this process across network, reducing interpersonal variability and ensuring treatment consistency across clinicians. Our study evaluated Siemens Healthineers' RT Image Suite VB 80 auto-segmentation algorithm for breast cancer target delineation modeled on RTOG guidelines across our network of cancer centers, aiming to assess its practical effectiveness. Material/Methods: A cohort of 54 female breast cancer patients who had previously undergone treatment was examined. Using AASH software, automated contours of breast, chest wall and nodal regions were generated. Ten physicians from various hospitals, within the network, participated in the study, evaluating the automated contours on a 4-point scale: a score of 4 indicated clinical usability without revisions, while scores of 3, 2, and 1 signified increasing need for minor edits, major edits, or complete re-contouring, respectively. Manual reviews of AI-generated contours were conducted by physicians, and analysis of scores was performed by an independent reviewer blinded to evaluators' identities. Data analysis was conducted using SPSS software. Results: The mean scores for Chest wall, Breast, Level 1 axilla, Level 2 axilla, Level 3 axilla, Supraclavicular and Internal Mammary lymph (IMN) nodes were 2.87(range 1.29-4), 2.97(2.5-3.95), 3.27(2.26-3.85), 3.54(2.94-3.95), 3.63(2.96 3.97), 3.61(2.96-3.97) and 3.34(2.44-3.96) respectively. 92.5% of chest wall, 70.6% of breast, 91.3% of supraclavicular, 87.4% of axilla 1, 96.7% of axilla 2, 97.6% of axilla 3, and 94.8% of IMN structures were graded as clinically usable with no or minor revision. There was significant variation in scoring between physicians. Chest wall had lowest score with commonest reason being non contiguity of contouring especially in regions of seroma and post-surgical changes. Conclusion: The AI-driven auto-segmentation software demonstrated proficient performance in delineating regional nodal target volumes as per RTOG atlas. It was usable with no or minor edits most of the time

Keywords: breast , nodes , segmentation