Author: Hao-Wen Cheng, Jonathan G. Li, Chihray Liu, Wen-Chih Tseng, Guanghua Yan 👨🔬
Affiliation: University of Florida 🌍
Purpose: This study develops and evaluates deep learning (DL) models for predicting 3D dose distributions in simultaneous integrated boost (SIB) prostate cancer treatment using the Elekta Unity MR-Linac. These models facilitate the quick determination of whether dosimetric criteria could be met for challenging treatment plans.
Methods: Six different DL architectures—standard U-Net, hierarchically dense U-Net, residual U-Net, cascaded U-Net, attention-gated U-Net, and dense dilated U-Net—were developed to predict 3D dose distributions for prostate cancer patients treated with SIB-intensity-modulated radiation therapy (SIB-IMRT) (60 Gy/44 Gy) on the MR-Linac. The model input consisted of a structure map combining 11 structures with their relative electron densities (REDs), two target masks with REDs, and an organs-at-risk (OARs) mask with REDs, all concatenated into a 3D image with 4 channels. The reference plans for 20 patients and their adaptive plans were included, resulting in 60 plans. Model performance was assessed using the mismatch criterion, which compares discrepancies in meeting dosimetric criteria for targets and OARs, as well as the dosimetric differences between the ground truth and the predictions.
Results: Most models showed small average dosimetric differences across the targets and OARs, with values typically within ±1.5 Gy and ±2% for most dosimetric parameters, indicating reasonable agreement with the ground truth. The average mismatch across different models was relatively low, ranging from 1.6 to 2.4 out of 20 dosimetric criteria. Among the evaluated models, the attention-gated U-Net model demonstrated superior performance, with the fewest mismatches and fewest instances of dosimetric differences > 0.5 Gy or 2%.
Conclusion: The attention-gated U-Net model effectively predicted the 3D dose distribution for SIB-IMRT prostate cancer treatment on the MR-Linac. This approach has the potential to streamline the treatment planning process by efficiently verifying whether dosimetric criteria are met, thereby reducing the substantial manual effort required in treatment planning.