Cross-Institutional Study of Geometric-Based Predictive Models for Cardiac Sparing in Breast VMAT: Free Breathing vs. Deep Inspiration Breath Hold

Purpose/Objective: Deep Inspiration Breath Hold (DIBH) has been demonstrated to reduce heart dose during breast radiotherapy. However, its benefits vary among patients, and clinicians currently lack a reliable tool to predict its effectiveness. This study explores the feasibility of utilizing Overlap Volume Histogram (OVH) parameters to develop predictive models for the heart dose in Volumetric Modulated Arc Therapy (VMAT) breast treatment under both DIBH and Free Breathing (FB) conditions in a multicentric context. Material/Methods: A total of 20 CT scans (10 FB and 10 DIBH) with corresponding RT structures DICOM files were randomly selected from a database of breast patients previously treated at Center A. VMAT plans were generated for the 20 CT series and structure set at two institutions using distinct technologies: Center A used Elekta Monaco 5.1 treatment planning system (TPS) with an Elekta Synergy Linac whereas Center B employed the Varian Eclipse TPS V11 with a TrueBeam Linac. A MATLAB 2024b (Mathworks) code was developed to compute the OVH, generating heart volume overlap thresholds at 10%, 20%, 30%, 40%, and 50% (denoted as R10%, R20%, R30%, R40%, and R50%, respectively). For each breathing technique, Single Center Models (SCM) were generated based on linear correlations between mean heart dose and coverage thresholds. Additionally, a Multicenter model (MCM) was developed using combined data from both centers. The evaluation of the quality of these models has been verified using Root Mean Square Error (RMSE) metric. Results: SCMs for each OVH parameter are summarized in Table1. High R- squared values and low RMSE values confirm the suitability of using these parameters for building predictive models. However, the analysis revealed distinct SCMs between FB and DIBH, with FB consistently showing steeper slopes compared to DIBH. Comparable correlations were observed between center A and center B. The MCM results, reported in Table2, indicate that the quality of the single-center and the multicenter models is comparable and confirm the necessity of using distinct models for the two breathing techniques. Conclusion: This study demonstrated the feasibility of building predictive models for both FB and DIBH techniques using OVH parameters. Furthermore, the findings suggest that data from multiple centers can be integrate into a single comprehensive model accommodating different technologies and planning approaches.