Source-linked AI summary
Dose-PlanNet: Physics Based Radiotherapy Dose Prediction with Deep Learning
Ankit Bhattacharjee, Sougata Maity, Santam Chakraborty, Indranil Mallick
TL;DR
Prostate SIB radiotherapy planning is complex because it must satisfy competing target-dose and organ-sparing constraints. Dose-PlanNet uses physics-guided 3D deep learning with spatial mappings and differentiable clinical penalties, achieving comparable target coverage and improved high-dose OAR sparing while meeting protocol criteria in many automated plans. The pipeline is intended to accelerate clinically viable planning, although inference required post-processing for ghost-dose artifacts.
Problem
Automating prostate radiotherapy planning requires complex dose distributions while maintaining strict target and organ-at-risk constraints.
Method
Dose-PlanNet combines discrete PTV mapping, spatial anatomy and beam-geometry inputs, and a multi-objective physics-guided loss with differentiable DVH penalties.
Results
Dose-PlanNet achieved comparable target coverage and superior high-dose organ-at-risk sparing relative to expert plans.
Takeaways & Limitations
The pipeline accelerates radiotherapy treatment planning while maintaining dosimetric quality for high-precision clinical deployment.
Takeaways & Limitations
Initial unconstrained inference produced approximately 5.0 Gy ghost-dose pockets in about 2 to 3 slices of some CT volumes, requiring a deterministic spatial bounding mask.
Abstract
from arXiv · showhide
Automating prostate radiotherapy treatment planning is dosimetrically complex, particularly for extreme hypofractionated regimens. In this study, we introduce Dose-PlanNet, a physics-guided 3D deep learning architecture designed to predict dose distributions. This model's performance was evaluated on a cohort of patients treated in a prospective trial where two different dose fractionation regimens were employed. Dose-PlanNet achieved comparable target coverage ($D_{95}$), though statistical analysis revealed a marginal reduction in target homogeneity ($p<0.001$) offset. However the model achieved statistically significant improvements in high-dose organ-at-risk sparing ($p<0.001$). When evaluated against strict Prospective Randomized protocol volumetric constraints, automated plans met prespecified clinical acceptance criteria in $11$ out of $14$ Moderate Hypofraction Arm plans and $9$ out of $12$ Stereotactic Body Radiation Therapy Arm plans. This pipeline demonstrates that physics-informed deep learning can accelerate radiotherapy workflows while safely maintaining the stringent dosimetric quality required for high-precision clinical deployment.
1 Introduction
Prostate SIB radiotherapy requires complex multi-target dose gradients while respecting strict anorectum and bladder constraints. Dose-PlanNet addresses limitations of voxel-wise prediction with physics-guided spatial and dosimetric optimization.
- SIB plans simultaneously deliver escalated primary PTV and lower elective-node doses, creating complex spatial gradients.
- Strict maximum and dose-volume constraints for the anorectum and bladder must be satisfied to prevent severe toxicities.
- 3D deep learning and KBP have enabled volumetric dose prediction and automated plan generation, but conventional methods rely mainly on MSE or MAE.
- Voxel-wise objectives can produce low global error while violating serial-OAR point-dose limits or creating cold spots at target peripheries.
- Dose-PlanNet combines discrete PTV mapping, Signed Distance Maps, a Beam’s Eye View mask, and a multi-objective physics-guided loss engine.
2 Related Work
Dose prediction research progressed from image-to-image translation toward anatomical embeddings and differentiable clinical constraints. Remaining gaps include inadequate optimization of strict boundaries, multi-objective SIB prescriptions, and physically deliverable beam geometry.
- Early frameworks used 3D U-Nets, FCNs, and GANs to map anatomy to volumetric dose or refine dose-cloud texture.
- Anatomical embeddings such as SDMs and distance-to-target coordinates improved geometric conformity but did not enforce strict clinical boundaries.
- Differentiable DVH metrics shifted optimization toward clinical acceptability, but most evaluations remained limited to single-target prescriptions.
- Constraint-aware loss engines can risk severe target underdosing without steep, dynamically weighted penalty gradients.
- Existing frameworks inadequately address SIB target overlaps and may deposit low-dose scatter in physically impossible out-of-beam regions.
- Dose-PlanNet addresses these gaps using Painter’s Algorithm PTV mapping and a BEV frustum mask to enforce dose hierarchies and beam-geometry constraints.
3 Methodology
Dose-PlanNet converts clinical treatment data into physics-aware spatial tensors and uses a 3D U-Net with progressively introduced clinical and geometric penalties to predict dose distributions. Its pipeline explicitly represents SIB prescription hierarchies, beam geometry, anatomical boundaries, and dose-falloff requirements during optimization.
- Data representation and preprocessing: The modular preprocessing pipeline aligns and digitizes patient anatomy, treatment geometries, and prescribed dose objectives from raw clinical DICOM data.
- Discrete SIB mapping: Painter’s Algorithm rasterizes target volumes by prescription dose so higher-dose targets overwrite overlapping lower-dose regions, preserving SIB dose hierarchy.This resolves ambiguity where elective and primary target volumes intersect.
- Beam geometry: BEV frustum masks are constructed from clinical gantry angles and ray-traced geometry to penalize dose deposited outside active beam corridors.The mask uses a 7.0 mm physical penumbra margin.
- Data representation and preprocessing: The network input combines CT, discrete PTVs, anatomical masks and SDMs, body information, penile bulb information, and a BEV frustum in a 7-channel tensor.The 3D U-Net maps this tensor to a continuous volumetric dose distribution.
- Physics-guided loss engine: Dose-PlanNet uses a custom multi-objective loss that balances voxel-wise fidelity with clinical boundary constraints and differentiable volumetric penalties.The loss includes spatial, geometric, and dosimetric components, with tuned coefficients balancing these objectives.
- Curriculum training: Physics-based and geometric penalties ramp linearly from zero to target magnitudes over the first 30 epochs, after which restrictive clinical penalties finalize conformity and OAR sparing.The warmup first stabilizes the global dose distribution through the constant dose-weighted reconstruction loss.
4 Experimental Setup and Evaluation
Dose-PlanNet was evaluated on two prostate radiotherapy cohorts spanning moderate hypofractionation and SBRT, using separate protocol-specific dosimetric objectives. Evaluation included held-out testing, an MSE-only ablation, and clinical review against institutional criteria and Eclipse-generated plans.
- Cohorts: 70 moderate-hypofractionation patients were split into 56 training and 14 held-out testing cases.
- Cohorts: 62 SBRT patients formed a separate cohort partitioned into training and held-out validation sets for hypofractionated dose prediction.
- Dosimetric criteria: D95 ≥95% target coverage was required in both arms, with maximum-dose limits of 66.34 Gy for moderate hypofractionation and 38.8 Gy for SBRT.
- Dosimetric criteria: Moderate-hypofractionation plans were assessed using bladder and anorectum V60.4Gy and V38.0Gy constraints plus femoral-head Dmax ≤40.0 Gy.
- Dosimetric criteria: SBRT plans were assessed using bladder and anorectum V35.0Gy and V31.5Gy constraints plus bilateral femoral-head V14.0Gy < 5%.
- Clinical evaluation: Clinical review was unblinded and qualitative, involving three consultants, two of whom were study co-authors; no structured rating instrument was used.
5 Implementation Details
The implementation uses a modular DICOM-to-dose pipeline with standardized spatial preprocessing and physics-informed inputs. Training combines normalized dose targets, targeted boundary sampling, scheduled physics penalties, and dual checkpoint selection for mathematical and clinical objectives.
- Pipeline architecture: A modular open-source Python stack with centralized configuration supports reproducibility and integration with clinical treatment planning systems.
- Preprocessing: A YAML configuration maps site-specific ROI names, while SimpleITK resamples volumes to a 1.27 × 1.27 × 2.5 mm3 grid and computes SDMs and BEV frustums.
- Dose normalization: Ground-truth dose tensors are normalized to [0, 1] by dividing by the primary prescription dose to stabilize training across prescription magnitudes.
- Sampling and optimization: Training patches of 128 × 128 × 64 voxels are anchored on intersections between target volumes and adjacent critical organs rather than sampled uniformly.
- Sampling and optimization: Adam, cosine-annealing learning rates, mixed precision, and curriculum learning ramp physics penalties from zero during initial epochs.
- Model selection: Dual checkpointing retains both the minimum-penalty model and a clinically optimized model using a softmargin minimax objective.
- Clinical deployment: Inference converts CT, RTSTRUCT, and RTPLAN DICOM inputs into predicted RTDOSE files and reverses spatial transformations to native CT geometry.
6 Results & Discussion
Dose-PlanNet produced clinically acceptable prostate radiotherapy dose predictions across moderate and extreme hypofractionation, with target coverage comparable to human planning and improved high-dose OAR sparing. Its trade-offs included marginally lower target coverage and isolated voxel-level Dmax artifacts, while maintaining clinical constraints.
- Ablation Study: The MSE-only ablation produced D95 = 33.5 Gy and Dmax exceeding 78 Gy against a 38.8 Gy clinical ceiling, demonstrating failure to satisfy spatial safety constraints.
- Ablation Study: Physics-guided penalties address MSE’s neglect of isolated spatial extremes and independently evaluated voxels, which can produce hotspots and non-deliverable dose gradients.
- Moderate Hypofraction Arm: D95 averaged 60.15 ± 0.74 Gy for PTV62, exceeding the mandatory 58.9 Gy threshold in the Moderate Hypofraction Arm.
- Moderate Hypofraction Arm: Eclipse achieved higher PTV62 D95 and lower Dmax, but both plans exceeded the mandatory target threshold and the model’s elevated Dmax was confined to isolated voxels.
- Moderate Hypofraction Arm: Dose-PlanNet reduced high-dose bladder and anorectum volumes versus Eclipse, with V59.0Gy values of 2.86 ± 1.11% and 3.48 ± 1.30%, respectively.
- Stereotactic Body Radiation Therapy Arm: In the SBRT Arm, PTV D95 averaged 34.89 ± 0.86 Gy while bladder and anorectum high-dose constraints were maintained at V35.0Gy values of 2.71% ± 0.92% and 4.24% ± 1.10%.
7 Limitations and Technical Refinements
Dose-PlanNet showed clinically acceptable target metrics and OAR sparing, but voxel-wise prediction produced artifacts and a small target-homogeneity trade-off requiring deterministic refinement.
- Dosimetric trade-offs: Slightly lower PTV D95 and marginally higher Dmax remained within mandatory clinical acceptability thresholds.The loss engine prioritized steep dose fall-off near critical structures, occasionally at the expense of absolute target homogeneity.
- Inference artifacts: Approximately 2–3 slices per CT volume contained isolated ∼5.0 Gy ghost-dose pockets distant from the primary isocenter.The artifact was associated with the expansive receptive field of the 3D U-Net.
- Inference artifacts: Low-dose wash regions occasionally showed discontinuous edges and isolated dose islands inconsistent with physical photon scatter.These irregularities could require physically impossible multileaf collimator sequencing for delivery.
- Technical refinements: A low-pass 3D Gaussian smoothing filter removed isolated islands and produced physically consistent, deliverable isodose contours.The refinement specifically addressed morphological irregularities in low-dose regions such as the 12.5 Gy contour.
8 Conclusion
Dose-PlanNet automated clinically viable prostate radiotherapy planning while matching target coverage and improving low-dose sparing for critical organs. Spatial bounding masks and Gaussian smoothing addressed voxel-wise artifacts, supporting protocol compliance and physical deliverability.
- Clinical conclusion: Dose-PlanNet achieved comparable target coverage and consistently superior low-dose sparing for the bladder and anorectum versus expert human plans.The evaluation used expert-generated plans from the Eclipse treatment planning system.
- Technical refinements: Single-voxel maximum-dose artifacts, ghost doses, and low-dose irregularities were neutralized with spatial bounding masks and 3D Gaussian smoothing.These refinements targeted limitations inherent to voxel-wise prediction.
- Clinical conclusion: Dose-PlanNet provides an autonomous pipeline intended to accelerate radiotherapy planning without compromising dosimetric quality.The proposed workflow is positioned for scalable, high-precision clinical deployment.
Data and Code Availability
The authors state that the Dose-PlanNet architecture, code, and supporting study data will be made openly available through the CHAVI-India GitHub repository.
- Availability: The architecture, relevant code, and supporting data are planned for open release in the CHAVI-India GitHub repository.The passage provides the repository URL: https://github.com/ CHAVI-India.
A Model Hyperparameters and Optimization Strategy
Both fractionation-arm models use a shared 3D U-Net training setup, with fixed manually tuned optimization parameters and efficiency-oriented training procedures.
- Hyperparameter strategy: Hyperparameters and loss weights λ were fixed after manual tuning through human evaluation of dosimetric boundaries across validation runs.The parameters were not dynamically scheduled.
- Model and training: Both models use a unified 3D U-Net backbone trained with PyTorch and MONAI.The same backbone is used for the Moderate Hypofraction and Stereotactic Body Radiation Therapy arms.
- Optimization: AdamW optimization is paired with a Cosine Annealing Learning Rate scheduler.This optimizer-scheduler combination is used for both fractionation-arm models.
- Training efficiency: Mixed-precision training with torch.cuda.amp accelerates gradient calculation and improves memory allocation.Mixed precision was used across all training iterations.
A.2 Physics-Guided Loss Engine Weightings
The physics-guided loss engine uses manually tuned λ weights to balance volumetric regression with physical and clinical constraints. Table 5 specifies distinct penalty configurations for the Moderate Hypofraction and SBRT arms.
- Weighting strategy: Manually tuned λ weights balance volumetric regression against physical boundaries and clinical mandates.The total loss combines these objectives through scalar coefficients.
- Protocol-specific settings: Table 5 specifies the λ coefficients used for the Moderate Hypofraction Arm and SBRT configurations.
- Protocol-specific settings: The two clinical protocols use distinct penalization choices mapped to their respective treatment requirements.
B Software Environment and Reproducibility
The study documents a software environment for external validation and computational reproducibility. The pipeline relies on core library versions to support volumetric regression and coordinate-space interpolation.
- Reproducibility: The implementation specifies core library versions to facilitate external validation and computational reproducibility.
- Software environment: The software environment is intended to maintain consistency in volumetric regression.
- Software environment: The documented environment also supports consistency in coordinate-space interpolation.