S2330
Physics - Quality assurance and auditing
ESTRO 2026
Digital Poster 562 Patient-Specific QA for Online Adaptive Prostate SBRT Plans CORAL LAOSA-BELLO 1 , SANDRA MORAGUES- FEMENIA 1 , ANTONIA PÉREZ-ZAMORA 1 , CLAUDIA LÓPEZ-ALARIO 1 , Juan-Francisco Calvo-Ortega 1,2 1 RADIATION ONCOLOGY, HOSPITAL QUIRÓNSALUD BARCELONA, BARCELONA, Spain. 2 RADIATION ONCOLOGY, HOSPITAL QUIRÓNSALUD MÁLAGA, MÁLAGA, Spain Purpose/Objective: Currently, patient-specific QA (PSQA) for online adaptive plans relies on a secondary independent dose calculation of the treatment plan. Retrospectively, we report the delivery accuracy of the type of online adaptive SBRT plans designed in our department for prostate treatment. Material/Methods: We perform prostate SBRT in five fractions using an in- house online adaptive method.1 To analyze the dosimetric accuracy of the delivered adaptive plans, the first 85 patients treated with a five fraction SBRT schedule (5 × 7.25 Gy) were included in this study. For each one, two adaptive plans were randomly chosen (170 adaptive plans were collected). Post-treatment experimental verifications of these clinical plans were performed using the PTW Octavius 4D system with the 1600 matrix (O4D16). Gamma passing rates (GPRs) were computed using two criteria: 1) the 3%(global)/2 mm criteria recommended by the AAPM TG-218 report,2 and 2) the stricter 2%(local)/2 mm gamma index criteria, both criteria using an exclusion threshold set to 10% of the maximum value. The AAPM TG-218 report establishes universal tolerance of GPR greater than 95%. Results: All plans passed the PSQA universal tolerance limit ≥ 95.0% despite the gamma analysis criteria. GPRs (mean ± SD) of 100% ± 0.0% (range: 99.8%-100%) and 99.1% ± 0.9% (range: 95.2%-100%) were found using the 3%(global)/2 mm and 2%(local)/2 mm criteria, Post-treatment measurements revealed that the deliverability of the adaptive plans met the universal tolerance limit ≥ 95% adviced by the AAPM TG-218 report for the 3%(global)/2 mm criteria. Besides, all plans passed this tolerence with the more demanding 2%(local)/2 mm criteria. These results, based on the retrospective analysis of a sample of 170 adaptive plans, ensure the dosimetric accuracy of our adaptive SBRT technique for prostate cancer. References: 1 Pract Radiat Oncol. 2022 Mar-Apr;12(2):e144- e1522Med Phys. 2018 Apr;45(4):e53-e83 respectively. Conclusion:
distributions achieving an overall accuracy of 0.96 and an ROC-AUC value of 0.94. SHAP-based explainability enabled identification of the most relevant features influencing gamma failures, facilitating clinical interpretation and corrective actions. Figure 1 illustrates the SHAP maps overlaid on the real and predicted portal and MM images, together with the corresponding feature contributions. Areas with higher SHAP values highlight regions and parameters that mostly influenced the model’s prediction discrepancies.
The SHAP analysis was prospectively applied to 10 patients whose initial portal verifications had failed (GPR 3%2mm < 90%). After implementing the corrective measures suggested by the explainability analysis and repeating the portal verification, an improvement in GPR was observed in 7 out of 10 cases (Figure 2), confirming the clinical relevance of the proposed approach.
Conclusion: This study demonstrates how SHAP interpretation using deep learning improves the prediction of full gamma maps in breast IMRT. Unlike scalar GPR prediction, the proposed approach identifies where and why verification failures occur, improving GPR results with specific corrective actions, bridging the gap between automated prediction and clinical interpretability. Further analyses are required to continue investigating the root causes of discrepancies and to achieve improved predictive performance. Keywords: explainability, deep learning, IMRT
Made with FlippingBook - Share PDF online