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; M. Gau; S. Stathakis; V. Grégoire; N. Paragios

cs.CVOct 19, 2026

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

Authors: J. Boyer-Chammard, R. Vauclin, M. C. Biston, P. Maury, M. Gau, S. Stathakis, V. Grégoire, N. Paragios,

Abstract

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.

Figures & tables

Figure 1: Overview of the proposed 3D latent diffusion framework for dose prediction. Patient and treatment information are encoded into volumetric latent representations that condition a 3D diffusion model. The generated dose latent is subsequently decoded into a voxel-space dose distribution.

Figure 2: End-to-end automated treatment planning framework.

Tumor primary site	Proportion (%)
Oropharynx	40
Hypopharynx/Larynx	30
Oral cavity	25
Others	5

Table 1: Composition of the in-house cohort (

n=900

): distribution of tumor primary sites (left) and of prescribed dose levels (right). Proportions are percentages of the full cohort and sum to 100% in each panel. Prescription groups refer to the dose delivered to the high-dose PTV.

Figure 3: Decodability of the perturbed latents, with or without denoising regularization (DenoReg). Each dose embedding is interpolated with Gaussian noise following Eq. ( 10 ) and decoded; columns show increasing corruption levels. Without regularization, reconstruction degrades beyond $\tau=0.2$ ; it remains stable until $\tau=0.6$ with denoising regularization.

Figure 4: Effect of latent denoising regularization (DenoReg) on the learned representations and latent space structure. Left: coronal view of a 3D dose distribution and the associated embeddings, with and without DenoReg. The regularized representation is smoother and its structures are more contrasted. Right: 2D t-SNE embedding of the training set in the latent space, computed after reduction to 150 principal components (perplexity 30) and colored by kernel density estimate, high density regions appearing in yellow.

	In-house test set		GDP-HMM
Metric	Baseline	DenoReg (Ours)	DenoReg (Ours)
MAE (Gy) $\downarrow$	$\mathbf{0.135}$ [1pt] $(0.120,\,0.153)$	$0.160$ [1pt] $(0.147,\,0.181)$	$0.312$ [1pt] $(0.27,\,0.36)$
DVH score (Gy) $\downarrow$	$0.160$ [1pt] $(0.139,\,0.185)$	$\mathbf{0.127}$ [1pt] $(0.107,\,0.180)$	$0.250$ [1pt] $(0.19,\,0.32)$
$\Delta$ HI (pp)	$-0.249^{\dagger}$ [1pt] $(-0.503,\,-0.077)$	$-0.292^{\dagger}$ [1pt] $(-0.625,\,-0.072)$	$+0.220^{\dagger}$ [1pt] $(-0.092,\,+0.652)$
$\Delta$ nCI	$-0.0139^{\dagger}$ [1pt] $(-0.0223,\,-0.0091)$	$+0.0026^{\dagger}$ [1pt] $(-0.0005,\,+0.0069)$	$-0.0242^{\dagger}$ [1pt] $(-0.0506,\,-0.0141)$
$\Delta D_{99}^{\mathrm{PTV}}$ (Gy)	$-0.139^{\dagger}$ [1pt] $(-0.182,\,-0.055)$	$\mathbf{+0.039}$ [1pt] $(-0.065,\,+0.154)$	$-0.381^{\dagger}$ [1pt] $(-0.736,\,+0.061)$

Table 2: Fidelity of the dose autoencoder. Reconstruction of the delivered dose through the encoder and decoder of Section 3.2 , under the protocol of Section 4.2 . The two in-house columns differ only in how the decoder was trained: without latent denoising, versus with latent corruption followed by fine-tuning on clean samples. The last column applies that configuration, without retraining, to the public GDP-HMM cohort. Dose and DVH scores are absolute errors;

\Delta

values are signed deviations from the delivered dose. Median

(Q_{1},\,Q_{3})

across patients; values in bold mark significant superiority, † marks a deviation significantly different from zero (two-sided Wilcoxon signed-rank,

\alpha=0.05

	Clinical	Predicted dose
Metric		DCNN	C3D	DoseDiff	MD-Dose	MedNeXt	LaDose
		Reg., 2D	Reg., 3D	Diff., 2D	Diff., 2D	Reg., 3D	Lat. diff., 3D
Volumetric accuracy
MAE (Gy)	– [1pt] $(0.000,\,0.000)$	$1.40^{*}$ [1pt] $(1.07,\,1.61)$	$1.97^{*}$ [1pt] $(1.73,\,2.35)$	$1.50^{*}$ [1pt] $(1.22,\,1.74)$	$2.18^{*}$ [1pt] $(1.81,\,2.88)$	$1.16$ [1pt] $(0.95,\,1.48)$	$1.18$ [1pt] $(0.90,\,1.44)$
DVH score (Gy)	– [1pt] $(0.000,\,0.000)$	$1.81^{*}$ [1pt] $(1.55,\,2.08)$	$1.64^{*}$ [1pt] $(1.36,\,2.11)$	$1.15$ [1pt] $(0.91,\,1.57)$	$1.83^{*}$ [1pt] $(1.36,\,2.84)$	$1.36^{*}$ [1pt] $(0.97,\,1.88)$	$1.19$ [1pt] $(0.91,\,1.57)$
Target coverage and dose shaping

Table 3: Evaluation of dose prediction performance on the test set. Dose and DVH scores are absolute errors; all other reported metrics are signed distances from the delivered clinical dose, with

\Delta=\text{predicted}-\text{clinical}

. The Clinical plan column gives the absolute value of the delivered plan for reference. We report median

(Q_{1},\,Q_{3})

across patients. Testing follows Section 4.2 : ∗ marks a significant difference from LaDose, † a significant bias with respect to the delivered plan.

Figure 5: Qualitative comparison on three test patients. Each patient is shown on axial, sagittal and coronal slices through the center of the high-dose PTV. Rows, from top: the delivered clinical dose, the dose predicted by LaDose, and their difference (LaDose $-$ delivered). Cases are the patients at the $25$ th, $50$ th and $75$ th percentile of the LaDose MAE distribution over the test set, so that they span the range of performance summarized in Table 3 rather than illustrating the best results. Prescriptions to the high-dose PTV are 70, 66 and 70 Gy. The high-dose PTV is outlined in white.

	Reference	Predicted dose
Metric	plan	MedNeXt	LaDose	$p_{\mathrm{Holm}}$
		Reg., 3D	Lat. diff., 3D
Volumetric accuracy
MAE (Gy)	– [1pt] $(0.000,\,0.000)$	$\mathbf{1.23}$ [1pt] $(0.91,\,1.45)$	$1.57$ [1pt] $(1.20,\,1.94)$	$1.5\!\times\!10^{-12}$
DVH score (Gy)	– [1pt] $(0.000,\,0.000)$	$1.03$ [1pt] $(0.78,\,1.32)$	$\mathbf{0.88}$ [1pt] $(0.70,\,1.03)$	$0.011$
Target coverage and dose shaping

Table 4: Comparison on the public AAPM GDP-HMM head-and-neck validation cohort, against the challenge-winning entry run from the authors’ released checkpoint. Dose and DVH scores are absolute errors; all other metrics are signed distances from the reference dose,

\Delta=\text{predicted}-\text{reference}

. Median

(Q_{1},\,Q_{3})

across patients. The last column gives the two-sided Wilcoxon signed-rank

p

between the two methods after Holm correction; † marks a significant bias with respect to the reference dose.

Components

Guidance

Error (Gy)

Target

OAR sparing (Gy)

DisCond

DenoReg

mCFG

s_{\mathrm{CT}}/s_{\mathrm{PTV}}/s_{\mathrm{OAR}}

MAE

\downarrow

DVH

\downarrow

\Delta

\Delta

nCI

\Delta D_{99}

\Delta

\Delta

\Delta

Par

\Delta

Mand

1.4/1.4/1.4

1.34

1.39

+1.05^{\dagger}

0.000

-0.11

-0.18

+1.34

-0.12

+0.33

✓

1.4/1.4/1.4

1.28

1.34

+0.88^{\dagger}

-0.004

+0.12

-0.37

+1.25

-0.56

+0.27

✓

1.4/1.4/1.4

\mathbf{1.18}^{*}

\mathbf{1.19}^{*}

\mathbf{-0.09}^{*}

-0.023

+0.07

+0.02

+0.67

-0.48^{*}

+0.03

✓

1.4/1.8/1.4

1.20

1.13

-0.22^{\ddagger}

-0.029^{\dagger\ddagger}

+0.33^{\ddagger}

-0.05

+0.54

-0.32^{\ddagger}

+0.11

✓

1.4/1.4/4.0

1.22

1.27

+0.38

-0.040^{\dagger}

-0.62^{\ddagger}

-0.41^{\ddagger}

-0.28^{\ddagger}

-0.59^{\dagger\ddagger}

-0.56^{\dagger\ddagger}

Table 5: Incremental ablation of the model components, on the same test set and under the protocol of Section 4.2 . Each cell is a cohort median over the test patients: the absolute error for MAE and the DVH score, and the signed deviation of the prediction from the delivered plan for the

\Delta

endpoints. Clinical reference values for the deviation endpoints are those of Table 3 . ∗ marks a significant difference from the preceding configuration, † a significant bias with respect to the delivered plan, ‡ a significant difference from the single-scale reference (third row). The last two rows share the checkpoint of the third and differ only in the inference-time guidance scales; they characterize controllability rather than accuracy, and are therefore not marked with ∗ .

Figure 6: Prediction performance on the test set (

n=76

) stratified by primary tumor site and prescription. Points are group medians, bars their

95\%

distribution-free confidence intervals (achieved coverage

96

–

99\%

, as such an interval can only fall on observed values). In each panel the dashed line and the gray band are the whole-cohort median and 95% confidence interval. The italic

p

is the Kruskal–Wallis test of heterogeneity across the subgroups for that metric.

Figure 7: Effect of guidance scales on dose quality metrics. Increasing the PTV guidance scale $s_{\mathrm{PTV}}$ improves the conformity of target coverage (nCI) at the expense of brainstem and parotid sparing. Increasing the OAR guidance scale $s_{\mathrm{OAR}}$ encourages spinal cord sparing. Solid lines show mean normalized values and shaded regions indicate $\pm 1$ standard deviation.

Structure	Metric	Clinical	Automated	$\Delta$ (95% CI)	$p_{\mathrm{Holm}}$
Organs at risk
Left parotid	$D_{\mathrm{mean}}$	$23.24\pm 13.31$	$20.16\pm 12.93$	$-2.43$ (-4.36—1.11)	$5.3\times 10^{-5}$
Right parotid	$D_{\mathrm{mean}}$	$24.03\pm 11.60$	$20.03\pm 11.15$	$-4.15$ (-5.41—2.66)	$5.3\times 10^{-5}$
Oral cavity	$D_{\mathrm{mean}}$	$36.28\pm 16.41$	$34.51\pm 16.12$	$-1.37$ (-3.50–+0.04)	$0.0482$
Spinal cord	$D_{2}$	$30.07\pm 5.28$	$25.75\pm 5.74$	$-3.79$ (-5.66—2.03)	$6.2\times 10^{-5}$
Brainstem	$D_{2}$	$15.00\pm 10.95$	$13.30\pm 9.77$	$-1.57$ (-2.79—0.24)	$0.0482$

Table 6: Dosimetric and plan-complexity comparison of automated and delivered clinical plans, for the structures contoured in essentially every patient. Values are mean

\pm

SD across patients;

\Delta

is the median within-patient difference (automated

-

clinical) with a 95% percentile confidence interval. Doses in Gy.

	Automated	Clinical	$p$
Acceptable as is, all reviewers	14/15 93% (70–99)	1/15 7% (1–30)	$2.9\times 10^{-6}$
Re-planning required, $\geq$ 1 reviewer	0/15 0% (0–20)	4/15 27% (11–52)	$0.10$
Complete reviewer agreement a	13/14 93% (69–99)	2/14 14% (4–40)	$6.9\times 10^{-5}$

Table 7: Clinical acceptability of automatically generated versus manually generated, clinically delivered plans. A plan is counted as acceptable only if every reviewer assigned to it rated it acceptable as is. Agreement is computed on the plans carrying two or more independent ratings. Proportions with 95% Wilson score intervals;

p

from two-sided Fisher exact tests.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Structure	Metric	$n$	Clinical	Automated	$\Delta$ (95% CI)	$q_{\mathrm{BH}}$
Delineated in $\geq$ 24/30 patients: tested
Larynx	$D_{\mathrm{mean}}$	25	$47.06\pm 18.35$	$44.91\pm 19.05$	$-2.10$ (-3.62—1.00)	$0.0507$
	$D_{5}$	25	$64.03\pm 13.79$	$60.87\pm 15.51$	$-2.32$ (-3.59—1.77)	$7.2\times 10^{-7}$
Esophagus	$D_{\mathrm{mean}}$	27	$10.50\pm 7.92$	$9.48\pm 6.62$	$-0.89$ (-1.59—0.02)	$0.0301$
	$D_{5}$	27	$31.92\pm 17.25$	$30.62\pm 17.15$	$-1.39$ (-2.51–+0.08)	$0.114$
Left submandibular gland	$D_{\mathrm{mean}}$	28	$49.00\pm 21.35$	$45.19\pm 21.13$	$-3.57$ (-4.95—2.70)	$3.2\times 10^{-4}$

Appendix

Table 8: Dosimetric comparison of automated and delivered clinical plans for the organs at risk and target volumes that were not delineated in every patient. Columns are as in Table 6 . Structures delineated in at least 24 of the 30 patients were tested with two-sided exact Wilcoxon signed-rank tests and the Benjamini–Hochberg procedure applied across that family; structures delineated in fewer patients are reported descriptively only, as the available sample does not support a meaningful test. Doses in Gy.

Endpoint	All $(n=76)$	Oropharynx $(n=31)$	Larynx $(n=14)$	Hypopharynx $(n=11)$	Oral cavity $(n=11)$	Other/unspecified $(n=9)$	KW $p$
MAE $\downarrow$ (Gy)	$1.18$ [1pt] $(0.90,\,1.44)$	$1.27$ [1pt] $(0.93,\,1.43)$	$1.27$ [1pt] $(0.99,\,1.46)$	$1.17$ [1pt] $(0.97,\,1.55)$	$1.16$ [1pt] $(0.77,\,1.35)$	$0.80$ [1pt] $(0.61,\,1.04)$	0.360
DVH $\downarrow$ (Gy)	$1.19$ [1pt] $(0.91,\,1.57)$	$1.30$ [1pt] $(0.93,\,1.54)$	$1.16$ [1pt] $(0.86,\,1.81)$	$1.34$ [1pt] $(1.06,\,1.59)$	$0.91$ [1pt] $(0.84,\,1.09)$	$1.29$ [1pt] $(0.97,\,1.63)$	0.359
$\Delta$ HI (pts)	$-0.09$ [1pt] $(-1.29,\,+1.24)$	$-0.20$ [1pt] $(-2.41,\,+0.52)$	$-0.02$ [1pt] $(-1.20,\,+1.77)$	$+0.53$ [1pt] $(-0.50,\,+1.14)$	$-0.12$ [1pt] $(-1.60,\,+1.09)$	$+0.78$ [1pt] $(-0.40,\,+1.98)$	0.417
$\Delta$ nCI	$-0.023$ [1pt] $(-0.069,\,+0.020)$	$-0.023$ [1pt] $(-0.057,\,+0.016)$	$+0.013$ [1pt] $(-0.117,\,+0.077)$	$-0.062$ [1pt] $(-0.098,\,-0.019)$	$-0.004$ [1pt] $(-0.049,\,+0.014)$	$-0.042$ [1pt] $(-0.111,\,+0.043)$	0.585
$\Delta D_{99}^{\mathrm{PTV}}$ (Gy)	$+0.07$ [1pt] $(-0.72,\,+1.12)$	$+0.22$ [1pt] $(-0.65,\,+1.13)$	$+0.05$ [1pt] $(-1.05,\,+0.40)$	$-0.47$ [1pt] $(-0.75,\,+0.21)$	$+0.10$ [1pt] $(-0.53,\,+1.29)$	$+0.12$ [1pt] $(-2.07,\,+1.60)$	0.848
$\Delta D_{0.1\mathrm{cc}}^{\mathrm{BS}}$ (Gy)	$+0.02$ [1pt] $(-1.86,\,+2.34)$	$-0.49$ [1pt] $(-1.61,\,+2.93)$	$+0.06$ [1pt] $(-0.84,\,+1.16)$	$-0.01$ [1pt] $(-5.24,\,+1.66)$	$+0.31$ [1pt] $(-2.24,\,+1.99)$	$+1.34$ [1pt] $(-1.76,\,+2.60)$	0.953

Appendix

Table 9: Prediction performance on the test set stratified by primary tumor site. Values are median

(Q_{1},\,Q_{3})

across patients. Dose and DVH scores are absolute errors; all other endpoints are signed deviations from the delivered clinical plan,

\Delta=\text{predicted}-\text{clinical}

. Structure abbreviations as in Table 3 .

Endpoint	All $(n=76)$	54–57.5 Gy $(n=6)$	60–64 Gy $(n=10)$	66 Gy $(n=15)$	69–70 Gy $(n=45)$	KW $p$
MAE $\downarrow$ (Gy)	$1.18$ [1pt] $(0.90,\,1.44)$	$0.78$ [1pt] $(0.59,\,0.85)$	$1.11$ [1pt] $(0.76,\,1.67)$	$1.16$ [1pt] $(0.77,\,1.36)$	$1.28$ [1pt] $(1.03,\,1.46)$	0.009
DVH $\downarrow$ (Gy)	$1.19$ [1pt] $(0.91,\,1.57)$	$1.43$ [1pt] $(1.10,\,1.71)$	$1.15$ [1pt] $(0.91,\,1.30)$	$1.19$ [1pt] $(0.93,\,1.59)$	$1.13$ [1pt] $(0.91,\,1.57)$	0.751
$\Delta$ HI (pts)	$-0.09$ [1pt] $(-1.29,\,+1.24)$	$+1.80$ [1pt] $(-1.17,\,+2.93)$	$+0.19$ [1pt] $(-1.15,\,+3.39)$	$-0.12$ [1pt] $(-2.33,\,+1.69)$	$-0.10$ [1pt] $(-1.21,\,+0.76)$	0.544
$\Delta$ nCI	$-0.023$ [1pt] $(-0.069,\,+0.020)$	$-0.032$ [1pt] $(-0.064,\,-0.013)$	$+0.060$ [1pt] $(-0.014,\,+0.100)$	$-0.023$ [1pt] $(-0.050,\,+0.017)$	$-0.033$ [1pt] $(-0.084,\,+0.006)$	0.075
$\Delta D_{99}^{\mathrm{PTV}}$ (Gy)	$+0.07$ [1pt] $(-0.72,\,+1.12)$	$-0.60$ [1pt] $(-1.61,\,+1.12)$	$+0.23$ [1pt] $(-0.86,\,+1.12)$	$-0.21$ [1pt] $(-0.50,\,+1.63)$	$+0.11$ [1pt] $(-0.65,\,+0.62)$	0.870
$\Delta D_{0.1\mathrm{cc}}^{\mathrm{BS}}$ (Gy)	$+0.02$ [1pt] $(-1.86,\,+2.34)$	$-0.37$ [1pt] $(-1.24,\,+2.04)$	$+0.30$ [1pt] $(-1.41,\,+1.40)$	$+0.64$ [1pt] $(-1.73,\,+3.42)$	$-0.23$ [1pt] $(-3.79,\,+2.39)$	0.915

Appendix

Table 10: Prediction performance on the test set stratified by the dose prescribed to the high-dose PTV. Columns and conventions as in Table 9 .

Radiotherapy Dose Prediction

See all 12 papers

May 13, 2026

Learning to Optimize Radiotherapy Plans via Fluence Maps Diffusion Model Generation and LSTM-based Optimization

Volumetric Modulated Arc Therapy (VMAT) is a cornerstone of modern radiation therapy, enabling highly conformal tumor irradiation and healthy-tissue sparing. Yet, its planning solves inverse and nested optimization for multi-leaf collimators, monitor units and dose parameters, while enforcing their consistency to ensure mechanical deliverability. Nevertheless, this process often requires repeated re-optimization when treatment configurations change, resulting in substantial planning time per patient. To address these problems, we present a diffusion-driven Learning-to-Optimize (L2O) method for end-to-end VMAT planning. A distribution-matching distilled diffusion model learns a clinically feasible manifold of fluence maps, enabling their one-shot generation. On top of this, an LSTM-based L2O module learns gradient update dynamics to swiftly refine fluence maps toward prescribed dose objectives during inference. Experimental results on clinical and public prostate cancer cohorts demonstrate improved planning efficiency, flexibility, and machine deliverability over currently available end-to-end VMAT planners.

Isabella Poles, Simon Arberet, Riqiang Gao +5

Politecnico di Milano, Milan, Italy · Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ, USA · Digital Technology and Innovation, Siemens Healthineers, Erlangen, Germany

Learning to Optimize Oncology Diffusion Models

Ask PDF

May 10, 2026

Any2Any 3D Diffusion Models with Knowledge Transfer: A Radiotherapy Planning Study

Voxel-wise dose prediction is a critical yet challenging task in practical radiotherapy (RT) planning, as bespoke models trained from scratch often struggle to generalize across diverse clinical settings. Meanwhile, generative models trained on billion-scale datasets from vision domains have achieved impressive performance. Herein, we propose DiffKT3D, a unified Any2Any 3D diffusion framework that leverages prior knowledge from pretrained video diffusion models for efficient and clinically meaningful dose prediction. To enable flexible conditioning across multiple clinical modalities (CT, anatomical structures, body, beam settings, etc.), we introduce an Any2Any conditional paradigm utilizing modality-specific embeddings without cross-attention overhead. Further, we design a novel reinforcement learning (RL) post-training mechanism guided by a clinically-informed Scorecard explicitly tailored to institutional treatment preferences. Compared with winner of GDP-HMM challenge, DiffKT3D sets a new state-of-the-art in dose prediction by reducing voxel-level MAE from 2.07 to 1.93. In addition, DiffKT3D achieves superior image quality and preference match. These results demonstrate that transferring diffusion priors via modality-aware conditioning and clinically aligned RL post-training can provide a robust and generalizable solution for RT planning across various clinical scenarios.

Yuhan Wang, Zihan Li, Han Liu +7

1UC Santa Cruz · 2Siemens Healthineers · University of Washington

Diffusion Model Alignment Radiotherapy Dose Prediction Conditional Diffusion Models

Ask PDF

Aug 27, 2026

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

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.

Ankit Bhattacharjee, Sougata Maity, Santam Chakraborty +1

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

Radiotherapy Dose Prediction Physics-Informed ML Oncology

Ask PDF

Abstract

Radiotherapy Dose Prediction

Explore similar work

Learning to Optimize Radiotherapy Plans via Fluence Maps Diffusion Model Generation and LSTM-based Optimization

Any2Any 3D Diffusion Models with Knowledge Transfer: A Radiotherapy Planning Study

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