Author: Yuzhen Ding, Hongying Feng, Jason Michael Holmes, Baoxin Li, Wei Liu, Daniel Ma, Lisa McGee, Samir H. Patel, Jean Claude M. Rwigema, Sujay A. Vora 👨🔬
Affiliation: Arizona State University, Department of Radiation Oncology, Mayo Clinic, Mayo Clinic Arizona, Mayo Clinic 🌍
Purpose:
Intensity-modulated proton therapy (IMPT) is a preferred treatment modality for head and neck (H&N) cancer patients, offering precise tumor targeting while sparing surrounding organs at risk (OARs). However, IMPT is highly sensitive to inter-fractional anatomical changes, necessitating periodic adjustments through online adaptive radiation therapy (oART). But a significant bottleneck in the current oART workflow is the generation of high-statistic influence matrices (IMs) during Monte Carlo (MC) simulations required for re-optimization. Using fewer particles (low-statistic IMs) can accelerate MC-based simulations but compromises accuracy. To address this, denoising low-statistic IMs is proposed as a method to rapidly and accurately produce high-statistic IMs.
Methods:
A diffusion transformer-based IM-denoising framework was developed. IMPT treatment plans from 10 H&N cancer patients were utilized, with high-statistic IMs and the corresponding low-statistic IMs derived by MC simulation, respectively. This dataset comprised 86 energy layers and 100,286 spots, with varying spot numbers per energy layer. Each data sample was vectorized into 1D vectors and tokenized into uniform chunks with zero-padding. The proposed method was first validated on single energy layer spot-wise denoising. For this, one spot per energy layer was reserved for testing, while the remaining data were used for training. The framework was then extended to multi-energy layers by training on data from 10 energy layers and predicting for the remaining energy layers. We embedded energy layer information as an additional input vector. Performance in both experiments was evaluated using mean-absolute-error (MAE), and profile comparisons were conducted to visualize differences between the predicted and ground-truth IMs.
Results:
The model achieved a MAE of less than 0.15% for single energy layer experiments and less than 0.5% for multi-energy layer experiments. IM profile comparisons also exhibited high agreement.
Conclusion:
A diffusion transformer-based framework was successfully developed, proving to be accurate in denoising low-statistic IMs.