Radiotherapy dose prediction using off-the-shelf segmentation networks: A feasibility study with GammaPod planning.

Abstract

Background: Radiotherapy requires precise, patient-specific treatment planning to achieve high-quality dose distributions that improve patient outcomes. Traditional manual planning is time-consuming and clinically impractical for performing necessary plan trade-off comparisons, including treatment modality selection, prescription dose settings, and organ at risk (OAR) constraints. A time-efficient dose prediction tool could accelerate the planning process by guiding clinical plan optimization and adjustments. While the deep convolutional neural networks (CNNs) are prominent in radiotherapy dose prediction tasks, most studies have attempted to customize network architectures for different diseases and treatment modalities.

Purpose: This study proposes a universal and efficient strategy, Seg2Dose, leveraging a state-of-the-art segmentation network for radiotherapy dose prediction without the need for model architecture modifications. We aim to provide a convenient off-the-shelf dose prediction tool that simplifies the dose prediction process, enhancing planning speed, and plan quality while minimizing the need for extensive coding and customization.

Methods: The proposed Seg2Dose consists of three modules: the Adapter, the segmentation network, and the Smoother. Prior to model training, the Adapter processes dose distributions into dose level map with an adjustable interval, which serves as the ground truth of the segmentation network, and generates two input channels: weighted avoidance image and normalized prescribed dose image. The segmentation network predicts dose levels from input channels using the nnU-Net, which was trained, validated and tested on 304, 77, and 64 breast cancer GammaPod treatment plans from 90 patients. The Smoother converts the predicted dose levels into continuous dose distribution with a Gaussian filter. The performance of Seg2Dose models with two different dose level intervals, 2% (Seg2Dose 2%) and 5% (Seg2Dose 5%), was evaluated by the Dice similarity coefficients (DSCs), voxel-based mean absolute percent error (MAPE), dose-volume histogram (DVH) metrics, global 3%/2 mm and 3%/1 mm gamma passing rate (GPR), and a case study including normal and worst cases. Additionally, Seg2Dose was compared with an exciting cutting-edge Cascade 3D (C3D) dose prediction model, which was trained on continuous dose distributions, to investigate the impact of using dose level map.

Results: For dose level prediction, Seg2Dose achieved average DSCs of 0.94 and 0.93 for the 2% and 5% intervals, respectively. For dose distribution prediction, both Seg2Dose 2% and Seg2Dose 5% achieved MAPEs within 6% for targets and most OARs, with the exception of the skin, which had the highest MAPE at 8.58% for Seg2Dose 2% and 15.25% for Seg2Dose 5%. The DVH metrics showed consistent findings. The C3D model has a better performance in GPR than Seg2Dose models. However, the C3D model exhibited higher MAPEs in target areas with lower dose predictions. In the case study, Seg2Dose 2% and C3D predictions were more consistent with clinical plans, showing smaller dose differences compared to Seg2Dose 5%.

Conclusions: Our study confirms the feasibility of leveraging the segmentation network for dose prediction and provides an efficient and off-the-shelf approach for dose prediction without requiring extensive coding efforts. This plug-in tool holds promise for quick dose planning, potentially aiding in the identification of optimal radiotherapy techniques and dosimetric tradeoffs prior to tedious treatment planning.