Enhancing treatment plan quality across the Mediterranean region using multicentric Overlap Volume Histograms (OVH) for left breast Radiotherapy

Purpose Knowledge-based planning (KBP) employs dose prediction models to automate and standardize radiotherapy treatment planning. Models developed using single-center data may be prone to bias due to suboptimal plan quality and limited patient anatomical diversity. The objective of this study was to conduct a multicenter investigation involving institutions from various geographical regions across the Mediterranean to evaluate quality improvement potentials through KBP for left breast radiotherapy. Comparisons were made between single-center and multicenter models, and model transferability across different centers was assessed. Methods Planning data of 160 left breast patients, all treated by tangential field 3D-CRT technique, were randomly selected from four centers (A, B, C, D) in Italy, Algeria, Libya, and Tunisia. The Overlap Volume Histogram (OVH, R10%-R50%) for heart and left lung were generated by custom MATLAB 2024b code, to build dose prediction models employing linear regression against dosimetric parameters (D2%, D10%, D20%, D50%, Dmean). Doses were normalized to PTV D98% to account for different treatment strategies across centers. Since the OVH-DVH relationship reflects plan quality in terms of OAR sparing, models were developed using high-quality plans from each center (single center models) and all centers combined (Multicentric model), establishing geometric-based dose benchmarks. For heart and left lung, each center's data was validated against its own single-center model, the other three centers' models, and the multicentric model. Data points exceeding the upper band of each model (±1σ) represented model uncertainty in predicting DVH based on OVH. Finally, all 160 patients' cohort was validated against the multicentric model following the same approach Results Figure 1 shows relevant inter-center variability for DVH prediction based on OVH. Table 1 reveals that internal validation detected 17-21 heart points outside the band (42.5%-52.5%) and 14-18 for lung (35.0%-45.0%) per center. External validation revealed center C's model detected fewest points outside the band (heart: 1-7, lung: 0-8) compared to other centers' data, while center D’s model detected most heart points outside the band (18-31). The multicentric model identified 16-27 heart points outside the band (40%-67.5%) and 7-32 for lung (17.5%-80%) across centers. Validating the complete 160-patient cohort against the multicentric model identified 81 patients (50.6%) and 68 patients (42.5%) outside the band for heart and left lung respectively Conclusion Our study showed that that single-center models have limited capacity to detect cases require further improvements. External and multicentric validations, incorporating broader geometric variability and treatment strategies, identified cases appearing acceptable within single institutions but deviating from multi-institutional standards