In radiotherapy, the goal is to deliver a high dose of radiation to the tumor while protecting surrounding healthy tissues like the heart and lungs. Calculating exactly how radiation will spread through a patient's body requires complex physics simulations that can take significant time.
Deep learning models that can predict dose distributions directly from CT scans have emerged as a faster alternative. However, breast cancer presents particular challenges because the breast is a soft tissue organ that varies greatly in size, shape, and density between patients.
Additionally, the radiation must avoid critical nearby structures - particularly the heart on the left side - meaning the spatial relationships between the tumor and surrounding anatomy are crucial to model accurately.
The researchers developed a 3D U-Net architecture - a network that processes full 3D CT volumes rather than 2D slices - to predict dose distributions across the entire treatment volume simultaneously. This captures the full spatial context of the patient anatomy.
Two novel modules were added to enhance the network. The Dense Encoder with Channel Attention (DECA) module helps the network focus on the most diagnostically relevant channels in the encoded features, while the Cross-Resolution Attention (CRA) module helps combine information across different spatial scales in the U-Net's skip connections.
The model was also trained with two custom loss functions: a 'difficulty perception loss' that gives more weight to regions where dose prediction is challenging, and a 'critical regions loss' that specifically penalizes errors in the target volume and organs at risk.
The model was trained and evaluated on a dataset of 176 breast cancer patients who had received radiotherapy. Each patient's CT scan, contoured structures (tumor and organs), and clinical treatment plan were used to create training examples.
The network learns to map from CT images and structure masks to the corresponding dose distribution. At test time, only the CT and contours are needed as input, and the network predicts where dose will be deposited.
Evaluation was performed using standard dosimetric metrics including the mean dose error, the dose coverage of the planning target volume, and the maximum dose to critical organs such as the heart, lungs, and opposite breast.
The model achieved a total dose error below 1.16 Gy across the treatment volume. Given that breast cancer patients typically receive 40-50 Gy total, this represents an error of around 2-3% - a level considered clinically acceptable for dose prediction tools.
The DECA and CRA attention modules both contributed to improved performance compared to a baseline 3D U-Net without these components. Ablation experiments confirmed each module's independent contribution.
The custom loss functions also improved results, particularly in the planning target volume and near the heart, where clinical accuracy is most critical. The difficulty perception loss helped the model focus training effort where errors mattered most.
The success of the attention modules suggests that not all spatial regions and feature channels are equally important for dose prediction. Learning to selectively emphasize the most relevant information mirrors how experienced planners focus their attention during manual plan evaluation.
Breast cancer cases present a wide range of anatomical configurations - left vs. right sided, different breast sizes, different cardiac positions. Attention mechanisms may help the model handle this variability more robustly than uniform processing.
The custom loss functions address a limitation of standard mean squared error training: treating all voxels equally regardless of clinical importance. By upweighting difficult and critical regions, the model is guided to perform where it matters most clinically.
This study presents a 3D U-Net model with novel attention and loss components that achieves clinically acceptable accuracy in predicting radiation dose distributions for breast cancer patients.
The modular design - where DECA, CRA, and custom losses can be independently evaluated and combined - provides a framework for continued improvement as more data becomes available and architectural innovations emerge.
Accurate dose prediction models like this one are a core building block for automated radiotherapy planning pipelines, enabling faster and more consistent treatment preparation across clinical sites.