Radiotherapy treatment planning requires a specialist to determine how to deliver radiation precisely to a tumor while minimizing exposure to surrounding healthy organs, a process that can take hours and requires expert judgment to balance competing dose objectives.
The goal of treatment planning is for the planning target volume (PTV), the tumor region, to receive the prescribed dose while organs at risk (OARs), such as the rectum, bladder, and femoral heads in prostate cases, receive as little radiation as possible. Even small reductions in OAR dose can meaningfully decrease side effects like rectal bleeding and urinary dysfunction.
Existing automated approaches include rule-based planning and knowledge-based planning (KBP), which predicts dose-volume histograms using regression methods like PCA or SVM trained on prior patients. However, these methods produce one-dimensional summaries and cannot capture the spatial arrangement of dose across the three-dimensional anatomy.
Deep learning offers a path to 3D voxel-level dose prediction, where every point in the treatment volume receives a predicted radiation dose value simultaneously, capturing the complex spatial gradients that are critical for evaluating plan quality and enabling fully automated treatment plan generation.
The proposed network, SA-Net (Scale Attention Network), is a 3D encoder-decoder architecture based on ResNet that processes entire volumetric CT images and organ contours simultaneously to predict a full 3D dose distribution map as output.
The key innovation is the scale attention (SA) block, which aggregates feature maps from all scales in the encoder rather than only using skip connections between matching scales as standard U-Net does. This allows the network to integrate fine anatomical details from shallow layers with high-level semantic patterns from deeper layers, learning to weight each scale's contribution adaptively based on the input case.
Each encoding and decoding block uses a ResSE block, a residual module enhanced with a squeeze-and-excitation mechanism that recalibrates the importance of each feature channel by compressing spatial information globally and learning channel-wise attention weights. This improves the network's ability to focus on the most informative feature channels for dose prediction.
SA-Net uses feature summation rather than concatenation between encoder and decoder pathways, reducing memory consumption while maintaining information flow. The total number of trainable parameters is 16.5 million, compared to 17.8 million for a comparable U-Net, making SA-Net more efficient while achieving superior performance.
Traditional dose prediction networks use binary masks as organ representations, where voxels inside an organ are assigned 1 and all others are assigned 0. This encoding provides no information about how far a voxel is from an organ surface, which is precisely the most important spatial relationship for predicting how much radiation a voxel will receive.
This study instead uses signed distance maps as input. For each organ, every voxel is assigned a value representing its Euclidean distance to the nearest point on that organ's surface in 3D space. Voxels inside the organ receive negative values (depth into the organ) and voxels outside receive positive values (distance from the surface).
The intuition is straightforward: radiation dose falls off as a function of distance from the tumor and rises as structures approach the radiation beam path. By providing the network with precise distance information for every organ, it can learn the geometric dose-falloff relationships that physically govern how radiation distributes through tissue.
Network inputs included signed distance maps for all contoured structures: the planning target volume (PTV), rectum, bladder, bladder wall, rectum wall, seminal vesicles, and femoral heads, plus the CT image. All data was standardized to a 256 x 144 x 64 voxel tensor at 3 x 3 x 3 mm resolution.
The study used a dataset of 140 prostate cancer patients treated with Volumetric Modulated Arc Therapy (VMAT) at Mount Sinai Hospital. VMAT is a sophisticated radiotherapy technique where a radiation beam rotates around the patient while continuously varying its shape and intensity, enabling highly conformal dose delivery to the prostate.
All clinical treatment plans used a prescription dose of 45 Gy. The actual clinically delivered dose distributions from these plans, calculated by the Eclipse treatment planning system (Varian Medical Systems), served as the ground truth labels for training the network.
The 140 patients were split into 100 for training, 20 for validation during training, and 20 for independent testing. The network was trained using the Adam optimizer with an initial learning rate of 0.0003 and ran for 60,000 iterations with batch size 2 on two NVIDIA GeForce 1080 Ti GPUs.
Mean Absolute Error (MAE) was used as the loss function, directly minimizing the average absolute dose difference between the predicted dose values and the clinical plan values at each voxel. This ensures the network optimizes for clinically relevant dose accuracy across the entire treatment volume.
On the 20-patient independent test set, the average absolute dose difference between the SA-Net predicted dose and the clinical treatment plan dose was only 0.94 Gy, equivalent to 2.1 percent of the 45 Gy prescription dose. This level of accuracy is within the typical clinical tolerance range for treatment planning acceptance.
For the planning target volume, the average predicted dose was 46.85 plus or minus 0.74 Gy, compared to the clinical planned dose of 46.81 plus or minus 0.78 Gy, a difference of less than 0.04 Gy on average. This near-perfect agreement confirms that the network learned to respect the prescription dose constraint for the tumor region.
Gamma analysis, a standard technique for evaluating dose distribution similarity that accounts for both spatial position and dose value differences, yielded an average gamma pass rate of 97.48 percent using a 5 percent dose difference threshold and 5 mm spatial distance criterion. This pass rate indicates that the predicted and clinical dose distributions are spatially and dosimetrically consistent across the vast majority of the treatment volume.
OAR dose constraints were also well preserved. For the rectum wall, the SA-Net predicted V30 (percentage of volume receiving 30 Gy or more) was 33.01 compared to the clinical plan's 36.73, and V40 was 16.72 compared to 21.84, suggesting that the network's predicted plans may even slightly reduce OAR exposure relative to standard clinical planning.
A direct comparison between using signed distance maps versus binary masks as organ representations confirmed the superiority of the distance map approach. The dose score (mean absolute error, lower is better) was 0.94 Gy with distance maps versus 0.99 Gy with binary masks, a 5 percent improvement.
The gamma pass rate also improved with distance maps: 97.48 percent versus 96.99 percent for binary masks. While the absolute difference may appear modest, in clinical radiotherapy even sub-percent improvements in dose accuracy can translate to meaningful reductions in treatment-related toxicity.
This comparison validates the core methodological hypothesis that providing explicit geometric distance information helps the network learn the physical relationship between organ proximity and dose distribution, rather than requiring it to infer that relationship from the abrupt 0-1 boundary provided by binary masks.
The finding has broad implications for the field of deep learning-based dose prediction, suggesting that preprocessing anatomy as distance fields rather than binary segmentation masks should become standard practice across all radiotherapy sites, not just prostate cancer.
SA-Net demonstrated strong voxel-level dose prediction for prostate cancer VMAT radiotherapy, with clinically acceptable accuracy and efficient use of GPU memory through its parameter-efficient architecture. The network's prior success in competitive medical imaging challenges (3rd place in MICCAI 2020 Brain Tumor Segmentation, 4th place in Head and Neck Tumor Segmentation) further validates its general capability for complex 3D medical predictions.
Automating dose prediction has direct clinical value. Current expert manual planning can take hours per patient. An automated system that instantly predicts a high-quality 3D dose distribution could serve as the starting point for planner refinement, dramatically reducing planning time and enabling more consistent plan quality across institutions with varying levels of planning expertise.
The combination of signed distance maps and scale attention creates a complementary system: distance maps provide physically grounded spatial context for each organ, while scale attention allows the network to integrate that context across multiple spatial resolutions, from fine edge details to coarse anatomical layout.
Future extensions could include applying the framework to other cancer sites such as head and neck, lung, and cervical cancer, adapting the model to different radiotherapy techniques such as IMRT and proton therapy, and developing real-time feedback systems that allow planners to interactively adjust constraints and immediately see the predicted dose response.