Author: Rafe A. McBeth, Kuancheng Wang, Ledi Wang 👨🔬
Affiliation: Department of Radiation Oncology, University of Pennsylvania, Georgia Institute of Technology, University of Pennsylvania 🌍
Purpose:
The integration of AI in clinical workflows presents unprecedented opportunities to enhance treatment quality in radiation oncology, yet it also demands innovative approaches to address the inherent complexity of clinical decision-making. Recent advancements in dose prediction have highlighted the potential of leveraging patient-specific anatomy, though the inherent tendency of deep learning models to converge toward average outcomes necessitates new strategies. This work builds on those insights by establishing a robust framework that incorporates clinically relevant metrics to refine prediction quality to accelerated partial breast irradiation (APBI) cases.
Methods:
Using the linear piecewise scoring framework, we assigned normalized scores to each treatment plan across a variety of dose-volume histogram (DVH) metrics, including target coverage and organ-at-risk (OAR) sparing criteria. To evaluate the distribution of plan quality across patient cohort, we applied the scoring framework to a retrospective dataset of 550 APBI patients treated in our clinic. We developed a deep learning model based on the Hierarchically Dense U-Net (HD U-Net) to predict the 3D dose distribution for APBI patients.
Results:
Preliminary results demonstrate significant variation in scoring distributions across the patient cohort, highlighting both patient-specific challenges and systemic trends in planning quality. Furthermore, we developed and evaluated a dose prediction model trained on a baseline dataset and an iteratively refined dataset guided by our dynamic scoring framework. The retraining process results in higher predicted plan quality, demonstrating the effectiveness of the refinement approach.
Conclusion:
This adaptive approach not only underscores the potential of AI to dynamically evolve with clinical needs but also highlights the critical role of scoring functions in guiding model refinement and clinical adoption. By addressing the inherent challenges of integrating AI in radiation oncology, this work lays the groundwork for the next generation of intelligent, high-precision treatment planning tools, offering improved outcomes for both clinicians and patients.