Radiation therapy is one of the leading treatments for cancer, and advances such as intensity-modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) have significantly improved patient outcomes over the past few decades.
However, these improvements come with a cost: greatly increased treatment plan complexity. A dosimetrist must go through many iterative adjustments and receive repeated feedback from a physician before a plan is approved, consuming significant clinical time.
Existing tools such as knowledge-based planning (KBP) have helped by predicting dose-volume histograms from historical patient data, but these methods rely on manually selected features and may not capture all the relevant information about a patient's anatomy.
A more powerful approach is to predict the entire dose distribution directly from a patient's anatomy -- essentially creating a map of exactly how radiation should be delivered -- which would give the clinical team a target to work toward from the very start of planning.
The researchers repurposed a deep learning architecture called U-net, originally developed for medical image segmentation, to instead predict radiation dose distributions from patient anatomy.
The model takes as input six contour images representing the planning target volume (PTV) and five organs at risk (OARs) -- bladder, body, left femoral head, right femoral head, and rectum -- and outputs a predicted dose map for each 2D slice of the patient.
Key architectural modifications included a seven-level hierarchy with six max-pooling operations to capture both fine local features and broad global patterns, plus seven additional convolutional neural network (CNN) layers to refine the prediction output.
To prevent the model from memorizing the training data (overfitting), the researchers implemented a dropout regularization scheme and used batch normalization after each activation function, finding this ordering significantly improved convergence speed.
The model was trained on 88 clinical prostate cancer patients, each treated with seven-field coplanar IMRT at 15 megavolts -- a relatively standardized treatment protocol well-suited for a feasibility study.
Because the central tumor-containing slices are most critical for dose accuracy, the team used a Gaussian sampling scheme that gave those central slices a higher probability of being selected during training, ensuring the model learned the most clinically important regions well.
A rigorous 10-fold cross-validation procedure was used to assess model performance on 80 patients, with a separate held-out test set of 8 patients to provide an unbiased final evaluation.
Performance was measured using several clinical metrics including Dice similarity coefficients for isodose volumes, mean and maximum dose differences for each structure, PTV coverage statistics, dose conformity, and dose spillage (R50).
The model achieved an average Dice similarity coefficient of 0.91 when comparing predicted versus actual isodose volumes across the full range from 0% to 100% of the prescription dose, indicating strong geometric agreement.
For individual structures, the average absolute dose prediction errors were well within 5% of the prescription dose: the PTV (tumor target) had errors of just 1.80% for maximum dose and 1.03% for mean dose.
Critical organs at risk were also predicted accurately: the rectum showed errors of 1.26% (max) and 1.62% (mean), the bladder 1.94% (max) and 4.22% (mean), and the femoral heads under 5.1% for all metrics.
PTV coverage statistics, dose conformity measured by the van't Riet conformation number, and dose spillage (R50) all showed less than 1% prediction error, demonstrating that clinically relevant plan quality metrics were well reproduced.
Visual inspection of dose predictions for patients with very different body geometries showed that the U-net had successfully learned to shape the predicted dose based on the size, location, and shape of the tumor target and surrounding organs.
The model captured fine-grained dose distribution details even in normal tissue regions far from the tumor, reflecting that it had learned meaningful relationships between anatomy and how radiation spreads through tissue.
One area of slightly lower performance was at the 40% isodose level, where the Dice coefficient dipped slightly -- attributable to complex dose patterns along beam paths in normal tissue, which are driven by the fluence map optimization process and are harder to predict from anatomy alone.
The current model was trained on patients treated with similar beam angles, meaning it may not generalize to more complex, non-standard non-coplanar beam geometries used in some treatment protocols.
The model also cannot currently incorporate physician preferences, such as requesting extra sparing of a specific organ like the rectum or bladder -- the prediction is driven purely by anatomy, not by clinical decision-making goals.
Future work aims to extend the approach to volumetric (3D) prediction models using architectures such as V-net, which could better handle non-coplanar plans that require understanding the full 3D patient geometry.
The researchers also plan to incorporate CT imaging data and dose constraint parameters as additional inputs, which would allow the model to personalize predictions based on tissue density and specific physician prescription requirements.
The immediate application of this technology is as a clinical guidance tool: by providing a predicted dose map at the start of planning, the dosimetrist and physician have a concrete reference that reduces the number of back-and-forth iterations required.
Since prostate IMRT protocols are largely standardized, this model could be deployed today in clinics that follow similar treatment protocols, offering a practical near-term benefit without waiting for further model development.
The long-term vision is an artificially intelligent treatment planning system capable of generating complete, clinically acceptable treatment plans with minimal human intervention, improving both efficiency and plan quality consistency across institutions.
By reducing dependence on individual planner experience and skill level, such systems could also help standardize plan quality across centers with different levels of expertise, making high-quality radiation therapy more accessible.