Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2554 - AI-Based Thoracic OAR Auto-Segmentation to Quantify Volumetric Differences Between 3D-CRT and IMRT in Whole-Breast Radiotherapy After Oncoplastic Reduction
Presenter(s)
Taran Singh, BS - New York Institute of Technology College of Osteopathic Medicine, Old Westbury, NY
J. Modi1, T. Singh1, K. Alabbasi2,3, I. Onochie1, L. Z. Braunstein1, D. A. Roth O’Brien2, A. J. Xu2, G. E. Naoum2, Z. Yehia2, J. Jiang4, H. Veeraraghavan4, A. Perez-Andujar1, L. I. Cervino4, I. J. Choi2,5, S. N. Powell2, A. J. Khan2, S. Elguindi4, and Q. LaPlant1; 1Memorial Sloan Kettering Cancer Center, New York, NY, 2Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 3Department of Physics and Astronomy, Hofstra University, Hempstead, NY, 4Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, 5New York Proton Center, New York, NY
Purpose/Objective(s):
Conventional whole-breast radiotherapy primarily evaluates lung and heart dose, while adjacent musculoskeletal thoracic structures are not routinely assessed due to contouring burden. In patients undergoing oncoplastic reduction, anatomic variability and tissue rearrangement may increase unintended irradiation of the chest wall and shoulder region. We hypothesized that AI-based auto-segmentation of novel thoracic OARs would enable scalable quantification of excess irradiated volume (EIV) and objective comparison of 3D-CRT and IMRT target volumes.Materials/Methods:
An in-house transformer-based foundation model was trained on 20 patients to auto-segment 51 bony and muscular thoracic structures and validated on 15 independent cases. Performance was graded (1 = no edits, 2 = minor edits, 3 = major edits, 4 = unacceptable) by a radiation oncologist and trained reviewers. The platform was then applied to 26 patients treated after oncoplastic reduction. For volumetric analysis, structures were grouped into composite regions (chest wall, pectoralis complex, and shoulder musculature [latissimus, teres major/minor, serratus anterior, and subscapularis]). AI-derived CTV and PTV_Eval generated predicted 3D and IMRT surrogates, enabling standardized comparison independent of clinical plans. EIV was defined as non-breast PTV tissue. AI and physician PTV_Eval agreement was assessed using Dice similarity coefficient (DICE) and mean distance to agreement (MDA).Results:
765 structures across 51 thoracic OAR classes in 15 patients were evaluated; all were identified. 99.87% were clinically usable and 98.8% required no or minor edits; one contour (0.13%) was unacceptable. Major edits were rare and limited to ribs. In 26 oncoplastic patients, AI- and physician-generated PTV_Eval volumes showed high agreement (Dice 0.906; MDA 2.59 mm). No edits were required for chest wall, pectoralis, or shoulder muscles. Breast volume strongly correlated with 3D EIV (r=0.88, p<0.0001). IMRT reduced EIV by ~70% versus 3D-CRT (median ratio 3.2), with reductions in every patient. Relative reductions were greatest in shoulder (87.8%), chest wall (79.7%), and pectoralis (58.2%), with the largest absolute reduction in chest wall (99.5 cc).Conclusion:
AI-enabled auto-segmentation of 51 thoracic OARs permits scalable, anatomy-aware quantification of planning inefficiency. Manual contouring would require hours per case, whereas automated generation is completed in minutes. In oncoplastic patients, IMRT demonstrates consistent and substantial reductions in excess irradiation of adjacent thoracic structures. This AI-assisted framework enables objective technique comparison and supports future outcome-correlated analyses.