Dose-PlanNet: Physics Based Radiotherapy Dose Prediction with Deep Learning

Ankit Bhattacharjee; Sougata Maity; Santam Chakraborty; Indranil Mallick

physics.med-phAug 27, 2026

Dose-PlanNet: Physics Based Radiotherapy Dose Prediction with Deep Learning

Authors: Ankit Bhattacharjee, Sougata Maity, Santam Chakraborty, Indranil Mallick

Organizations: Department of Mechanical Engineering, Indian Institute of Technology Kharagpur · Department of Radiation Oncology, Tata Medical Center Kolkata

Abstract

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.

Figures & tables

Figure 1: End-to-End Architecture of the Dose-PlanNet Pipeline. Raw clinical DICOM files are isotropically resampled and converted into physics-aware spatial embeddings, including Signed Distance Maps (SDMs), Beam’s Eye View (BEV) frustums, and a discrete Simultaneous Integrated Boost (SIB) map via Painter’s algorithm. The 7-channel tensor feeds a 3D U-Net trained via a multi-objective physics loss engine, where clinical constraints are progressively introduced during early training. The pipeline yields dual checkpoints (Physics-Optimized and Clinically-Optimized), which are utilized during sliding-window inference. Crucially, the final step mathematically reverses the affine resampling operations, restoring the predicted dose matrix to the native CT geometry for immediate Treatment Planning System (TPS) integration.

Figure 2: Decomposition of the Physics-Guided Loss Engine. The network optimizes a custom multi-objective loss function ( $\mathcal{L}_{total}$ ) driven by complex spatial embeddings. Constraints are categorized into Absolute Dose penalties (D-Type, targeting specific dose thresholds like PTV coverage and anti-collapse), Volume-Histogram penalties (V-Type, utilizing a differentiable sigmoid estimator for OAR sparing), and Spatial/Regularization penalties (enforcing geometric falloff and scatter suppression). A curriculum scheduler dynamically ramps the magnitude of these physics penalties from zero over the initial 30 epochs to ensure early numerical stability.

Type	ROI	Metric	Eclipse TPS (Mean $\pm$ SD)	Dose-PlanNet (Mean $\pm$ SD)	Mandatory Goal
Target	PTV62	$D_{95}$ (Gy)	61.39 $\pm$ 0.83	60.15 $\pm$ 0.74	$\geq 58.90$
	PTV62	$D_{max}$ (Gy)	65.74 $\pm$ 0.66	66.66 $\pm$ 0.68 ∗	$\leq 66.34$
	PTV44	$D_{95}$ (Gy)	43.54 $\pm$ 0.32	43.57 $\pm$ 0.60	$\geq 41.80$
Avoidance	Bladder	$V_{59.0\text{Gy}}$ (%)	3.82 $\pm$ 1.45	2.86 $\pm$ 1.11	$\leq 3.0$
	Bladder	$V_{40.0\text{Gy}}$ (%)	15.65 $\pm$ 5.24	14.25 $\pm$ 4.93	$\leq 55.0$
	Anorectum	$V_{59.0\text{Gy}}$ (%)	4.87 $\pm$ 0.93	3.48 $\pm$ 1.30	$\leq 3.0$

Table 1: Dosimetric Comparison: Dose-PlanNet vs. Human-Generated Eclipse TPS Plans on the Prostate Prospective Randomized Moderate Hypofraction Arm (62.0 Gy Prescription,

n=14

*Statistically significant at $\alpha=0.05$ .
Metric	Eclipse Mean	Model Mean	Mean Diff.	95% CI	p -value
PTV $D_{95}$ (Gy)	61.39 $\pm$ 0.83	60.15 $\pm$ 0.74	-1.24	[-1.84, -0.63]	0.0007*
Bladder $V_{59\text{Gy}}$ (%)	3.83 $\pm$ 1.47	2.87 $\pm$ 1.12	-0.96	[-1.39, -0.53]	0.0003*
Bladder $V_{40\text{Gy}}$ (%)	15.66 $\pm$ 5.25	14.26 $\pm$ 4.94	-1.40	[-2.66, -0.14]	0.0323*
Anorectum $V_{59\text{Gy}}$ (%)	4.87 $\pm$ 0.93	3.48 $\pm$ 1.30	-1.40	[-2.05, -0.75]	0.0004*
Anorectum $V_{40\text{Gy}}$ (%)	21.35 $\pm$ 5.52	20.50 $\pm$ 6.49	-0.85	[-2.87, 1.18]	0.3835

Table 2: Statistical comparison of dosimetric metrics between Eclipse TPS and Dose-PlanNet.

Type	ROI	Metric	Eclipse TPS (Mean $\pm$ SD)	Dose-PlanNet (Mean $\pm$ SD)	Mandatory Goal
Target	PTV36.25	$D_{95}$ (Gy)	35.46 $\pm$ 0.48	34.89 $\pm$ 0.86	$\geq 34.44$
	PTV36.25	$D_{max}$ (Gy)	38.57 $\pm$ 0.14	39.49 $\pm$ 0.63 ∗	$\leq 38.79$
	PTV25	$D_{95}$ (Gy)	25.70 $\pm$ 3.12	26.01 $\pm$ 2.81	$\geq 23.75$
Avoidance	Bladder	$V_{35.0\text{Gy}}$ (%)	2.91 $\pm$ 0.80	2.71 $\pm$ 0.92	$\leq 5.0$
	Bladder	$V_{14.0\text{Gy}}$ (%)	31.14 $\pm$ 7.04	22.62 $\pm$ 4.65	$\leq 50.0$
	Anorectum	$V_{35.0\text{Gy}}$ (%)	3.42 $\pm$ 1.26	4.24 $\pm$ 1.10	$\leq 5.0$

Table 3: Dosimetric Comparison: Dose-PlanNet vs. Human-Generated Eclipse TPS Plans on the Prostate Prospective Randomized Stereotactic Body Radiation Therapy Arm (36.25 Gy Prescription,

n=12

Panel of Figure 3 — Figure 3: Axial spatial dose distribution and isodose conformity comparison for a representative extreme hypofractionated SBRT case. (a) The human-generated Eclipse TPS plan and (b) the Dose-PlanNet prediction are overlaid on the native planning CT. Isodose contours denote 36.25 Gy (100% prescription, red), 34.44 Gy ( $D_{95}$ clinical mandate, orange), 25.00 Gy (cyan), and 12.50 Gy (low-dose wash, blue). The model demonstrates highly conformal high-dose target wrapping and robust spatial avoidance of adjacent critical structures, mirroring the clinical baseline.

Panel of Figure 4 — Figure 4: Comparative Dose-Volume Histograms (DVHs) for two representative patients. The solid lines correspond to the Dose-PlanNet (model-generated) dose, while the dashed lines correspond to the Eclipse TPS (human-planned ground truth) dose. Target volumes (red/blue curves) show strong conformity at the prescription shoulder, while OARs (green/magenta curves) demonstrate a steeper low-dose fall-off in the model predictions.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Moderate Hypofraction Arm	SBRT Arm
Maximum Epochs	500	300
Warm-up Epochs	30	30
Base Learning Rate	$1.0\times 10^{-4}$	$1.0\times 10^{-4}$
Minimum Learning Rate	$1.0\times 10^{-6}$	$1.0\times 10^{-6}$
Batch Size	1	1
Gradient Accumulation Steps	2	2

Appendix

Table 4: Training and Optimization Hyperparameters.

Reconstruction Fidelity
Loss Component	Notation	Moderate Hypofraction Arm $\lambda$	SBRT Arm $\lambda$
Dose-Weighted MSE	$\lambda_{mse}$	25.0	25.0
Anti-Collapse Penalty	$\lambda_{anticollapse}$	150.0	150.0
D-Type Target Penalties
PTV Coverage Hinge	$\lambda_{ptv}$	75.0	75.0
PTV Maximum Dose Hinge	$\lambda_{ptv\_max}$	30.0	30.0

Appendix

Table 5: Physics-Guided Loss Engine Penalties (

\lambda

Weights).

Library	Version
Python	3.10.2
PyTorch	2.3.0+cu121
MONAI	1.3.0
SimpleITK	2.3.1
PyDICOM	2.4.4
SciPy	1.13.0

Appendix

Table 6: Core dependencies and library versions.

Radiotherapy Dose Prediction

See all 12 papers

Learning the Dose Manifold: A Latent Diffusion Framework for Volumetric Dose Prediction in Automated Radiotherapy Treatment Planning
Oct 19, 2026J. Boyer-Chammard, R. Vauclin, M. C. Biston +7Radiotherapy Dose Prediction 3D Diffusion Models
AI End-to-End Radiation Treatment Planning Under One Second
Mar 6, 2026Simon Arberet, Riqiang Gao, Martin Kraus +9Oncology
Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ, USA. · Digital Technology and Innovation, Siemens Healthineers, Erlangen, Germany. · Varian Medical Affairs, Siemens Healthineers, Palo Alto, CA, USA.
A Machine-to-Machine Knowledge-Guided LLM Agent for Generalizable Radiotherapy Treatment Planning
May 30, 2026Md Mainul Abrar, Xun Jia, Yujie ChiLLM Planning
Department of Physics, The University of Texas at Arlington, Arlington, 76019, TX · Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Baltimore, MD

Oct 19, 2026J/K move · Enter open · S save

Learning the Dose Manifold: A Latent Diffusion Framework for Volumetric Dose Prediction in Automated Radiotherapy Treatment Planning

J. Boyer-Chammard, R. Vauclin, M. C. Biston, P. Maury+6

Radiotherapy Dose Prediction 3D Diffusion Models Latent Diffusion Models

Open paper AskPDF arXiv

Automated treatment planning (ATP) aims to reduce the unwarranted inter-planner variability of manual planning, while improving treatment quality and efficiency. Most ATP pipelines rely on the prediction of the achievable three-dimensional dose distribution for a given anatomy and the subsequent optimization of machine parameters to reproduce it. Diffusion-based methods are largely slice-wise due to the high dimensionality of radiotherapy data, and assume a one-to-one mapping between anatomy and optimal dose despite the multi-objective nature of treatment planning. They are evaluated on their ability to match reference doses and rarely on their integration in ATP frameworks. We present LaDose, a latent diffusion framework for fully volumetric dose prediction. An autoencoder learned on dose distributions, trained so that perturbed latents decode to plausible doses, compresses dose maps 128-fold and enables 3D diffusion on a single GPU. Prescription values are represented as exact scalar vectors that bypass the image encoder, avoiding resampling and latent compression. Multiscale classifier-free guidance allows for the modulation of clinical trade-offs at inference, without retraining or additional labeled data. On 900 clinically delivered head-and-neck volumetric modulated arc therapy plans from five centers, LaDose reached a median absolute error of 1.18 Gy (1.4--1.8% of the prescription) and was the only one of six frameworks free of systematic bias with respect to the reference distribution. Guidance scales acted predictably on target coverage and organ-at-risk sparing. Integrated with direct aperture optimization, the pipeline generated plans in 15.2 minutes; in a blinded review, senior clinicians rated 14 of 15 plans acceptable without editing.

Abstract

Radiotherapy Dose Prediction

Explore similar work

Learning the Dose Manifold: A Latent Diffusion Framework for Volumetric Dose Prediction in Automated Radiotherapy Treatment Planning

AI End-to-End Radiation Treatment Planning Under One Second

A Machine-to-Machine Knowledge-Guided LLM Agent for Generalizable Radiotherapy Treatment Planning