Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2608 - Development and Multi-Physician Validation of AI-Based Segmentation of Breast CTV and Tangent Field Geometry In Whole-Breast Radiotherapy
Presenter(s)
Taran Singh, BS - New York Institute of Technology College of Osteopathic Medicine, Old Westbury, NY
T. Singh1, J. Modi1, J. Pagan2, L. Z. Braunstein1, D. A. Roth O’Brien3, A. J. Xu3, F. Chino4, J. Jiang2, H. Veeraraghavan2, A. Perez-Andujar1, L. Hong2, I. J. Choi3,5, L. I. Cervino2, S. N. Powell3, A. J. Khan3, S. Elguindi2, and Q. LaPlant1; 1Memorial Sloan Kettering Cancer Center, New York, NY, 2Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, 3Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 4Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 5New York Proton Center, New York, NY
Purpose/Objective(s):
Thoracic and breast anatomy exhibit substantial inter-patient variability that directly influences 3D conformal tangent field geometry in whole-breast radiotherapy. In this paradigm, physicians must balance optimal breast coverage—influenced by lumpectomy cavity location and patient-specific anatomy—against adjacent normal tissue exposure. We sought to develop and multi-physician validate an AI-based model capable of generating breast CTV and derived PTV_Eval volumes that capture anatomy-dependent physician-intended tangent field geometry.Materials/Methods:
A cohort of 55 physician-defined whole-breast cases from two attending radiation oncologists, intentionally representing a broad range of breast volumes and tangent geometries, was used to fine-tune an in-house transformer-based foundation AI model for breast CTV and a derived PTV_Eval structure. PTV_Eval was designed to approximate the 95% isodose volume from 3D conformal tangents and was defined as 0.5 cm from the posterior field edge (excluding heart and lung) and 0.5 cm from the skin anteriorly. Model performance for breast CTV and PTV_Eval were evaluated on an independent test cohort (n=23). Agreement between AI-generated and physician-defined contours was assessed using mean distance to agreement (MDA) and Dice similarity coefficient (DICE). To assess robustness in routine clinical practice, the derived PTV_Eval model was applied to 108 consecutive whole-breast cases planned with 3D conformal tangents across 16 attending radiation oncologists. Statistical comparisons were performed between the independent validation cohort and attending-level real-world cases.Results:
On the independent test cohort (n=23), mean CTV DICE was 0.921 (95% CI 0.907–0.935) and mean MDA was 1.87 mm (95% CI 1.65–2.09). For the derived PTV_Eval structures, mean DICE was 0.916 (95% CI 0.907–0.926) and mean MDA was 1.898 mm (95% CI 1.665–2.131). Across 108 real-world PTV_Eval cases, mean DICE was 0.898 (95% CI 0.892–0.905) and mean MDA was 2.55 mm (95% CI 2.331–2.763). Performance was consistent across laterality and attendings (attending-level mean DICE 0.855–0.930; MDA 1.38–3.77 mm). No statistically significant differences in DICE or MDA were observed between the independent validation cohort and attending-level real-world cases.Conclusion:
An AI-based model achieved high geometric agreement for automated breast CTV and derived PTV_Eval segmentation across an independent test cohort and 108 consecutive real-world cases spanning 16 attendings. By capturing anatomy-dependent physician-intended tangent geometry, this framework supports standardization of target definition in 3D conformal whole-breast radiotherapy and enables prospective integration of AI-generated tangent geometry to streamline planning workflows, potentially expediting transition from simulation to treatment planning with focused physician review.