Source-linked AI summary
Three-Dimensional Radiotherapy Dose Prediction on Head and Neck Cancer Patients with a Hierarchically Densely Connected U-net Deep Learning Architecture
Dan Nguyen, Xun Jia, David Sher, Mu-Han Lin, Zohaib Iqbal, Hui Liu, Steve Jiang
TL;DR
Head-and-neck radiotherapy planning is complex and time-consuming, motivating automated volumetric dose prediction. The proposed Hierarchically Densely Connected U-net outperformed Standard U-net and DenseNet in dose-prediction quality while using fewer parameters and predicting faster.
Problem
Head-and-neck radiotherapy planning involves complex targets, multiple prescriptions, and nearby organs at risk, requiring extensive expertise and potentially up to one week.
Method
The study develops a Hierarchically Densely Connected U-net that combines U-net’s spatial information with DenseNet’s feature propagation for volumetric dose prediction.
Results
HD U-net outperformed Standard U-net and DenseNet in homogeneity, conformity, and coverage, predicting OAR Dmax within 6.3% and Dmean within 5.1% of prescription dose.
Takeaways & Limitations
The model can support clinical guidance by identifying organs at risk that may require greater attention during treatment planning.
Takeaways & Limitations
The study’s relatively small dataset may limit generalization and predictive performance.
Abstract
from arXiv · showhide
The treatment planning process for patients with head and neck (H&N) cancer is regarded as one of the most complicated due to large target volume, multiple prescription dose levels, and many radiation-sensitive critical structures near the target. Treatment planning for this site requires a high level of human expertise and a tremendous amount of effort to produce personalized high quality plans, taking as long as a week, which deteriorates the chances of tumor control and patient survival. To solve this problem, we propose to investigate a deep learning-based dose prediction model, Hierarchically Densely Connected U-net, based on two highly popular network architectures: U-net and DenseNet. We find that this new architecture is able to accurately and efficiently predict the dose distribution, outperforming the other two models, the Standard U-net and DenseNet, in homogeneity, dose conformity, and dose coverage on the test data. Averaging across all organs at risk, our proposed model is capable of predicting the organ-at-risk max dose within 6.3% and mean dose within 5.1% of the prescription dose on the test data. The other models, the Standard U-net and DenseNet, performed worse, having an averaged organ-at-risk max dose prediction error of 8.2% and 9.3%, respectively, and averaged mean dose prediction error of 6.4% and 6.8%, respectively. In addition, our proposed model used 12 times less trainable parameters than the Standard U-net, and predicted the patient dose 4 times faster than DenseNet.
I. INTRODUCTION
H&N radiotherapy planning is highly complex and labor-intensive, requiring iterative coordination between planners and physicians. This motivates the Hierarchically Densely Connected U-net, which combines U-net and DenseNet principles while reducing memory demands.
- H&N radiotherapy planning is complicated by large planning target volumes, multiple dose levels, and nearby radiation-sensitive critical structures.
- Treatment planners iteratively tune inverse-optimization hyperparameters and weightings through trial and error because physician-preferred plans are largely unknown.
- U-net is a convolutional architecture designed to incorporate local and global features for semantic segmentation.
- DenseNet densely connects convolutional maps to promote feature propagation and reuse, reduce vanishing gradients, and decrease trainable parameters.
- The proposed Hierarchically Densely Connected U-net combines U-net and DenseNet principles while requiring less memory than a 3D densely connected U-net.
II. METHODS · II.1. H&N Patient Data
The study used data from 120 head-and-neck cancer patients, including structure contours and clinically delivered VMAT dose distributions. Inputs represented OARs and PTVs in separate channels at 5 mm3 voxel resolution, without including patient CT.
- II.1. H&N Patient Data: 120 H&N patients were acquired for the study.The cohort included diverse H&N cancer sites, including tongue, tonsil, larynx, pharynx, thyroid, and other regions.
- II.1. H&N Patient Data: Table 1 summarizes the patient information.
- II.1. H&N Patient Data: Each patient contributed structure contours and clinically delivered VMAT dose calculated using the Eclipse Treatment Planning System.
- II.1. H&N Patient Data: 5 mm3 was the voxel resolution for both structure contours and dose.
- II.1. H&N Patient Data: Each OAR was encoded as a separate binary-mask input channel.Each voxel received 1 when assigned to the OAR and 0 otherwise.
- II.1. H&N Patient Data: Patient CT was not included as an input, while PTVs were included as their own channel.
II.2. Deep Learning Terminology
This section defines core deep learning concepts used in the paper, including CNN-based feature extraction, activation functions, pooling, and loss functions. These concepts describe how networks extract features, transform representations, reduce resolution, and measure prediction error.
- Convolutional Neural Networks: CNNs use kernels and convolution to extract image features such as edges, while providing shift-invariant extraction with fewer weights than fully connected networks.CNNs were introduced for computer vision and imaging tasks.
- Activation Functions: ReLU is a nonlinear activation function defined as f(x) = max(0, x), valued for strong performance, computational efficiency, and avoiding vanishing-gradient issues.It is typically applied after convolutional or fully connected operations.
- Pooling: 2 x 2 patches for 2D inputs and 2 x 2 x 2 patches for 3D inputs, with stride 2, halve each output dimension while retaining important features.Pooling aggregates patch information into a lower-resolution feature map, reducing computational expense and supporting more global image processing.
- Loss Functions: Loss functions quantify a model state as a single value that training optimizes by changing network parameters.In supervised deep learning, loss typically measures error between predictions and known ground truth using the network’s weights.
II.3. Deep Learning Architectures
The study implements three-dimensional HD U-net, standard U-net, and DenseNet architectures for volumetric head-and-neck data. HD U-net combines dense convolution and downsampling with U-net upsampling, using specified growth, downsampling, and upsampling settings.
- Hierarchically Dense U-net: HD U-net combines dense convolution, dense downsampling, and U-net upsampling operations.Dense convolution applies ReLU convolution followed by concatenation of the previous feature set, while U-net upsampling follows the standard U-net operation.
- Hierarchically Dense U-net: 16 new features are added after each dense operation, with 4 dense downsampling operations and 64 features returned during upsampling.
- Comparison Architectures: The standard U-net matches HD U-net in downsampling-operation count and doubles filters after each max-pooling operation.It uses regular convolution and max pooling for comparison with HD U-net.
- Comparison Architectures: 7 dense blocks, 5 dense convolutions per block, and a 0.5 compression factor define the DenseNet configuration matched approximately to HD U-net’s trainable parameters.DenseNet uses dense convolutions within dense blocks followed by compression layers.
- Implementation: 3D operations are used in all networks to process volumetric head-and-neck data.Dropout and L1/L2 regularization were excluded after training and validation curves indicated no overfitting; batch normalization was also removed from the final models.
II.4. Training and Evaluation
Training and evaluation used a held-out test set, 5-fold cross-validation, and standardized dose normalization. Models were trained with mean squared error and assessed using target coverage, dose, homogeneity, conformity, and organ-at-risk dose metrics.
- Data Splitting and Cross-Validation: 20 patients were held out for final testing, while the remaining 100 underwent 5-fold cross-validation with 80 training and 20 validation patients per fold.Weights were randomly initialized during each fold.
- Training Procedure: Mean squared error between predicted and clinically delivered dose trained each model, with learning rates adjusted to minimize validation loss across epochs.Training used 96 x 96 x 64 patches and a batch size of 1 sample.
- Dose Normalization and Evaluation: 95% of the highest-prescription PTV volume was required to receive 100% of the prescription dose (D95) for plan normalization.Dose statistics and errors were reported relative to prescription dose.
- Dose Normalization and Evaluation: Evaluation criteria included PTV coverage (D98, D99), PTV max dose, homogeneity, conformity, and structure max and mean doses (Dmax and Dmean).The conformity criterion used VPTV × V100%Iso.
III. RESULTS
The HD U-net generalized better than the Standard U-net and produced lower dose-prediction errors than the other models. It also showed superior test-set PTV dose coverage and performed favorably in example-patient dose predictions.
- Training and validation loss: Lower validation loss and fold-to-fold variation indicated that the HD U-net generalized better and overfit less than the Standard U-net.The HD and Standard U-net had similar training loss, but the HD U-net had much lower validation loss with less cross-validation variation.
- Computational performance: 12 times less trainable parameters were used by both the HD U-net and DenseNet than by the Standard U-net.The HD U-net required about 3 hours to train, compared with 2.4 hours for the Standard U-net and 8.8 hours for DenseNet.
- PTV coverage and dose: Statistically significantly better test-set dose coverage was achieved by the HD U-net than by the Standard U-net.The models had similar D95, D98, and D99 performance on cross-validation data.
- OAR and PTV dose errors: 6.23 ± 1.94% was the HD U-net’s average Dmax prediction error on cross-validation data, versus 8.11 ± 1.87% for the Standard U-net and 7.65 ± 1.67% for DenseNet.For Dmean, the corresponding errors were 5.30 ± 1.79%, 6.38 ± 2.01%, and 6.49 ± 1.43%, respectively.
- Dose prediction quality: The HD U-net had the lowest overall Dmax and Dmean prediction errors among the three networks.For an example test patient, it was also superior in predicting dose to the PTVs, although OAR predictions were more variable.
- Dose visualization: Comparable PTV dose predictions were observed visually, while the Standard U-net and DenseNet were slightly hotter than the HD U-net.DenseNet predicted dose above 3.5 Gy everywhere in the body, and all models predicted more dose to the back of the neck than the ground truth.
IV. DISCUSSION
The study presents an accurate volumetric dose-prediction approach for H&N cancer patients treated with VMAT, while acknowledging comparison, data, memory, and input limitations. The model is positioned as clinical guidance rather than autonomous treatment planning, with future work targeting improved inputs and AI-based planning.
- Novelty: This is presented as the first accurate volumetric dose prediction for H&N cancer patients treated with VMAT.Existing prediction approaches are largely based on Knowledge Based Planning, including RapidPlan and Pinnacle Auto-Planning Software.
- Limitations: Direct comparison with non-commercial prediction models is difficult because their proprietary implementations and institutional training data are generally unavailable.The authors also plan a comprehensive comparison against Varian RapidPlan, which uses geometrical relationships and volume features but not relationships among OARs themselves.
- Architecture: 64 slices may be required for accurate dose prediction at complex H&N sites, making prior hybrid dense U-nets infeasible on modern GPUs.The study motivates a less RAM-intensive densely connected U-net to address this constraint.
- Limitations and future work: The study’s data size is a limitation, although the authors did not observe overfitting in the training and validation curves.Larger datasets may further improve the model, while future work will evaluate physician-prescribed dose constraints and CT images as additional inputs.
- Clinical application: 6.3% and 5.1% are the HD U-net’s average OAR Dmax and Dmean prediction errors relative to prescription dose.The tool is intended for clinical guidance, while physicians and dosimetrists retain final trade-off and plan-delivery decisions.
- Future directions: The long-term goal is to transition from historical-plan dose prediction to an AI-based treatment-planning system that improves on current or past clinical practice.The authors intend to continue investigating new deep-learning concepts and architectures for more intelligent and efficient models.
V. CONCLUSION
The study developed HD U-net for volumetric dose prediction in head and neck cancer patients. Using PTV and OAR contours with prescription dose, the model accurately predicted dose distributions and OAR dose metrics on test data.
- Model and application: HD U-net was developed and applied to volumetric dose prediction for patients with head and neck cancer.The architecture is hierarchically densely connected and U-net-based.
- Model and application: PTV and OAR contours plus prescription dose served as inputs for predicting the dose distribution.The implementation predicts dose from these anatomical contours and prescription information.
- Test-data results: 6.3% was the average OAR max dose prediction deviation from prescription dose, while 5.1% was the average OAR mean dose deviation on test data.Both metrics are reported relative to the prescription dose.
APPENDIX · A.1. Details on deep learning architectures used in study
The appendix specifies the notation and common convolutional components used to describe the study’s deep learning architectures.
- A.1. Details on deep learning architectures used in study: All convolutions use 3 x 3 x 3 kernels followed by ReLU activation, while Dense Conv and U-net Upsample follow Figure 1 notation.These conventions define the architectural details summarized in Table 7.