Background: Determining the optimal energy layer (EL) for each field, under considering both dose constraints and delivery efficiency, is crucial to promoting the development of proton arc therapy (PAT) technology.
Purpose: This study aimed to explore the feasibility and potential clinical benefits of utilizing machine learning (ML) technique to automatically select EL for each field in PAT plans of lung cancer.
Methods: Proton Bragg peak position (BPP) was employed to characterize EL. The ground truth BPPs for each field were determined using the modified ELO-SPAT framework. Features in geometric, water-equivalent thicknesses (WET) and beamlet were defined and extracted. By analyzing the relationship between the extracted features and ground truth, a polynomial regression model with L2-norm regularization (Ridge regression) was constructed and trained. The performance of the regression model was reported as an error between the predictions and the ground truth. Besides, the predictions were used to make PAT plans (PAT_PRED). These plans were compared with those using the ground truth BPPs (PAT_TRUTH) and the mid-WET of the target volumes (PAT_MID) in terms of relative biological effectiveness-weighted dose (RWD) distributions. One hundred ten patients with lung cancer, a total of 7920 samples, were enrolled retrospectively, with 5940 cases randomly selected as the training set and the remaining 1980 cases as the testing set. Nine patients (648 samples) were collected additionally to evaluate the regression model in terms of plan quality and robustness.
Results: With regard to the prediction errors, the root mean squared errors and mean absolute errors between the ML-predicted and ground truth BPPs for the testing set were 9.165 and 6.572 mm, respectively, indicating differences of approximately two to three ELs. As for plan quality, the PAT_TRUTH and PAT_PRED plans performed similarly in terms of plan robustness, target coverage and organs at risk (OARs) protection, with differences smaller than 0.5 Gy(RBE). This trend was also observed for dose conformity and uniformity. The PAT_MID plans produced the lowest robustness index and lowest doses to OARs, along with the highest heterogeneity index, indicating better protection for OARs, improved plan robustness, but compromised dose homogeneity. Additionally, for relatively small tumor sizes, the PAT_MID plan demonstrated a notably poor dose conformity index.
Conclusions: Within this cohort under investigation, our study demonstrated the feasibility of using ML technique to predict ELs for each field, offering a fast (within 2 s) and memory-efficient reduced way to select ELs for PAT plan.
Keywords: energy layer prediction; lung cancer; machine learning; spot‐scanning proton arc therapy.
© 2024 American Association of Physicists in Medicine.