An interactive virtual dosimetrist for adjusting radiotherapy dose quality

Abstract

Background and Purpose:

Effective education in radiotherapy plan quality review requires exposure to treatment plans of varying quality and opportunities for plan improvements. However, the current clinic-based paradigm does not fully support these needs. Trainees often review limited numbers of diverse treatment plans and are not fully exposed to the planning decisions essential for developing strong evaluation skills.

Materials and Methods:

To address these challenges, we developed “Virtual Dosimetrist” models that generate and modify radiotherapy dose distributions through language conditioning, similar to communicating with a planner during plan review. This enables creation of diverse suboptimal examples and real-time, language-guided refinement.

Results:

The models modify dose distributions following language directives, allowing the generation of lower-quality examples and subsequent improvement. Dose predictions were produced quickly and without reliance on a planning system, enabling interactive adjustment that mirrors the clinical review-edit cycle. For the 11 patients in the test set, good control over the extent and magnitude of dose changes across 17 OOIs, together with robustness to variations in the language directives, was observed. When comparing to dose originating from a TPS, median RMSE was 52 cGy across all dose-level comparisons and 98% of predictions met the prespecified equivalence criterion of RMSE $<$ 200 cGy.

Conclusions:

This work is the first to combine deep learning-based dose prediction with natural language processing specifically designed to support the language-conditioned generation of suboptimal dose distributions and their improvement. The framework could be used as a foundation for educational applications, such as replicating plan quality review in an interactive environment.

Graphical abstract

Highlights

• •Clinic-based training alone makes radiotherapy plan quality education challenging.
• •“Virtual Dosimetrist” models adjust dose in response to language directives.
• •Examples of low- or improved-quality dose can be generated with language directives.
• •Approach can support future education applications like plan review training.

1. Introduction

Radiation oncology residents report substantial gaps in radiotherapy plan quality assessment and review skills [1]. Clinic-based training may contribute, as residents encounter a limited range of disease sites and typically review only finalized plans, not the many possible lower-quality variations. In addition, because radiotherapy planning is manual and time-intensive, substantial delays occur between identifying potential problems and receiving edits. This makes it difficult for trainees to understand which potential changes are most beneficial. Together, these challenges may impact patient care, with propagation of suboptimal plans through review and treatment well-documented [2], [3], [4].

Available training data could be expanded through multi- institutional plan curation. However, institutional differences limit its benefit, intermediate plans are rarely saved, and repeated viewing risks memorization over skill development. The second challenge, delayed feedback, is particularly difficult to address, as there are fundamental limitations on how quickly plan review iterations can be conducted in a typical clinic.

Recently, deep learning dose prediction has shown strong results across diverse treatment sites, particularly with knowledge-based approaches [5], [6], [7], [8], [9]. However, these models typically produce only high-quality (e.g., Pareto-optimal) dose distributions. These do not generate the range of lower-quality examples necessary for education.

To address this, we developed “Virtual Dosimetrist” dose modification models. These cross-modal models require only basic patient data (CT scan, initial dose, structure masks) and a natural language prompt describing the desired change. They generate varying-quality dose distributions in real time, enabling suboptimal examples and interactive modifications mirroring clinical edit-review processes. To our knowledge, this is the first text-conditioned, deep learning-only approach for degrading and improving dose quality to support educational applications.

2. Materials and methods

2.1. Data generation, curation, and augmentation

Fifty-three head-and-neck (HN) volumetric-modulated arc therapy (VMAT) physician-approved treatment plans were curated from our clinical practice. The MD Anderson Cancer Center Institutional Review Board approved this retrospective study, with protected health information removed and informed consent waived per institutional policy. The HN site was chosen for its complexity, as successful design of a virtual dosimetrist increases confidence in applicability to less-complex disease sites. High-dose target prescriptions ranged from 6000–7000 cGy, most commonly 6996 cGy, and all plans contained 1–5 additional lower-dose targets.

Suboptimal treatment plans were generated in a research RayStation (RaySearch Laboratories, Stockholm) version 12A treatment planning system (TPS) to have poor organ-of-interest (OOI) sparing by intentionally increasing OOI dose. For each OOI with defined planning objectives, auxiliary structures directly contributing to its optimization (e.g., expansions or false structures) were identified. Because clinical plans were highly optimized, relaxing planning objectives did not reliably increase dose. Instead, each OOI and auxiliary structures were temporarily reset to “target” type so the optimizer prioritized achieving dose goals, after which planning objectives for these structures only were iteratively adjusted to increase OOI dose. All other planning objectives were unmodified. This was applied individually to each OOI grouped with its auxiliary structures; dosimetrically or anatomically related structures (e.g., bilateral OOIs) were processed separately.

This was repeated for five iterations per OOI group to create progressively worse plan quality examples. Each iteration’s calculated dose was exported for training and served as the subsequent iteration’s starting point. TPS dose calculations used the collapsed cone convolution technique with other clinical plan parameters (e.g., beams, arcs, structure delineations) left unmodified. Dose was normalized post-optimization to preserve target coverage. A line search update of planning objectives based on prior OOI mean dose changes guided each step. This was repeated for OOI maximum dose changes to diversify dose change patterns. Finally, the planning CT, clinical and replanned doses, target structures, and OOI structures were exported with a dose grid spacing of 3 mm $\times$ 3 mm $\times$ 3 mm.

Distributions exceeding 8400 cGy were excluded, resulting in 4072 final dose distributions (Table 1) and avoiding clinically-unrealistic plans relative to the typical 6996 cGy prescription. The final dataset contained 18–140 suboptimal plans (median: 76) per clinical VMAT plan. An 80%/20% patient-level split yielded 42 training patients (3214 doses) and 11 test patients (858 doses). Five-fold cross-validation trained five models on separate data folds to improve generalization.

Table 1.

Number of dose distributions per OOI in the final dataset.

OOI (Left/Right)	# Dose distributions
Parotid (L)	420
Parotid (R)	403
Esophagus	373
Brainstem	357
Larynx	355
Spinal Cord	348
Cochlea (L)	320
Cochlea (R)	310
Lungs	248
Oral Cavity	238
Brachial Plexus (R)	142
Brachial Plexus (L)	125
Submandibular Gland (L)	111
Mandible	102
Submandibular Gland (R)	100
Lens (L)	60
Lens (R)	60

During training, CT Hounsfield units were clipped to (−200, 450) and normalized to (0, 1). Dose maps were normalized by 8400, and target masks were multiplied by prescription dose and normalized. OOI masks were isotropically dilated by 1.5 cm. Dose feature/label pairs used clinical/replanned (67%) or replanned/replanned (33%) combinations. Samples were equally likely to prompt a dose increase or decrease. Inputs were randomly cropped to (96, 112, 176) in (z, y, x). Random rotation (±3°), zoom (±10%), and x-y plane flip were each applied with 20% probability. Training prompts followed

• •Increase the metric dose to OOI name by amount cGy
• •Decrease the metric dose to OOI name by -amount cGy

where metric is minimum, mean, or maximum dose. Requested changes were strictly positive for increase prompts and strictly negative for decrease prompts.

2.2. Architecture

A dual-encoder architecture [10] was designed for this study (Fig. 1). Volumetric inputs (CT, dose map, potential dose mask (e.g., voxels where dose could be non-zero), target mask, and dilated OOI mask) were processed with a 3D DDU-Net, selected for its robust dose prediction performance [7]. Language prompts describing the desired dose modification were encoded using the OpenAI CLIP text encoder, chosen for its medical-language performance and cross-modal capabilities [11], [12], [13]. Publicly-available OpenAI CLIP weights were used, with additional multi-layer perceptrons (MLPs). The architecture was written in PyTorch 2.5.1 and trained on an Nvidia A40 GPU.

Fig. 1.

Overview of model architecture. Unless otherwise specified, all convolution kernel dimensions are 3 × 3 × 3, with the number of channels indicated in the figure.

All convolution operations except the final used batch normalization [14] and rectified linear unit (ReLU) activations, with the final convolution followed by ReLU only to avoid shifting dose values. Models were trained with batch size 8 using the Adam [15] optimizer. Initial learning rate was $8.5\times 10^{-5}$, following previously described scaling strategies [16], [17], and decreased tenfold after 50 epochs without validation loss improvements. Training proceeded for up to 1000 epochs with early stopping after 120 epochs of no reduction in validation loss.

A weighted mean-squared error (MSE) loss was computed separately for the dose mask, target mask, and dilated OOI mask, with the OOI region weighted most heavily to prioritize accuracy where dose changes were requested. The loss is shown in Eq. (1), where dm, PTV, and OOI denote the dose mask, target mask, and dilated OOI mask, respectively, and $D$ and $\hat{D}$ denote the ground truth and predicted dose map,

$$\mathcal{L}=0.2MSE(D_{dm},{\hat{D}}_{dm})+0.2MSE(D_{PTV},{\hat{D}}_{PTV})+0.6MSE(D_{OOI},{\hat{D}}_{OOI})$$ (1)

where MSE is defined in (2).

$$MSE(D,\hat{D})=\frac{1}{n}\sum _{i=1}^n{(D_i-{\hat{D}}_i)}^2$$ (2)

2.3. Inference and post processing techniques

For full-volume inference, a sliding-window approach was adapted from MONAI [18]. Patches overlapped by 50%, with Gaussian weighting emphasizing central patch contributions. An iterative approach was developed to improve overall dose accuracy (Supplementary Material). The final dose output was the unweighted ensemble average of the five cross-validation models. Similar to Gronberg et al. [8], normalizing the dose distributions at each iteration by the ratio of median dose to the high-dose target (3) preserved reasonable target coverage.

$$D_{\text{final}}=D_{\text{predicted}}\cdot \frac{D_{50,\text{initial}}}{D_{50,\text{predicted}}}$$ (3)

Models were trained to produce changes to minimum, mean, or maximum dose. Because voxel-wise extrema are noise-sensitive, surrogates were used for analysis. While prompts specified minimum/ maximum dose, results report dose to 99% of the volume (D99%) and to 0.1 cubic centimeters (D0.1cc) as robust surrogates for minimum and maximum dose, respectively. Mean dose was reported unmodified.

Selected TRIPOD+AI [19] principles applicable to this generative dose-modification study were observed. Additionally, the CLAIM [20] checklist was followed to ensure transparent reporting for this architecture.

2.4. Prompt robustness

Since the models were trained on relatively templated prompts, a prompt robustness experiment was conducted for both increase and decrease dose prompts (Supplementary Material).

2.5. Equivalence and error testing

For each TPS-generated suboptimal plan in the test set, dose change from the clinical plan was measured for minimum, mean, and maximum metrics. Across 11 test-set patients, the 858 plans yielded 2574 dose-change data points (3 per plan).

To assess model outputs to TPS-generaged dose, models were prompted to introduce these same changes into the clinical dose. Voxel-wise equivalence and error between predicted and TPS distributions were calculated in cGy and assessed with root mean square error (RMSE). Equivalence was defined a priori as RMSE $<$ 200 cGy, and additionally assessed with a two one-sided t test (TOST) with 100 cGy equivalence bounds ($p<0.0001$), consistent with prior studies [8], [9], [21]. To avoid bias, calculations excluded voxels outside the potential dose mask.

Because multiple modified plans were produced per patient, individual plan-level observations were not statistically independent. Therefore, results were analyzed in two ways. First, an OOI-stratified plan-level analysis summarized equivalence by structure. For clinical context, differences between predicted and TPS distributions for OOI mean dose and D0.1cc and PTV D95% and V100% were calculated for each OOI change. Second, a patient-level analysis summarized equivalence per patient to address clustering, reporting RMSE with 95% bias-corrected and accelerated bootstrap confidence intervals (95% CI) using 10,000 resamples.

2.6. Code availability

Institutional rules do not permit sharing the custom codebase and model weights. The implementation used standard open-source software (e.g., PyTorch 2.5.1), and the architecture description provides additional design details. The open-source code and weights for CLIP are provided by OpenAI within a software repository at https://github.com/openai/CLIP.

3. Results

The virtual dosimetrist models successfully generated dose distributions reflecting requested OOI dose increases and decreases. After selecting an OOI mask, the user prompt follows the training prompt format. A complete pass through the iterative refinement inference pipeline takes an average of 7 s to generate new dose distributions on an Nvidia A40 GPU.

For 500 cGy increase prompts, models produced dose changes consistent with the requested modification across D99% (near-minimum), mean, and D0.1cc (near-maximum) metrics (Table 2). Dose-decrease prompts showed similar trends. Mean-dose control was most consistent across modifications, whereas near-minimum and near-maximum response varied by structure. Models exhibited robustness to varying increase and decrease prompts (Supplementary Material Table S1). Fig. 2 shows an example of the model-generated dose changes. A suboptimal example was created from a clinical plan by prompting an increased mean esophageal dose, then improved by prompting a decrease, generating a distribution similar to the clinical plan.

Table 2.

Mean changes in cGy between predicted and clinical dose to specified dose metric, following the specified prompt to increase or decrease dose. $\sigma$: standard deviation.

Prompt	“Increase the metric dose by 500 cGy”			“Decrease the metric dose by −500 cGy”
Prompt metric	Minimum	Mean	Maximum	Minimum	Mean	Maximum
Evaluation metric	$\Delta D99\text{\%}$ ($\sigma$)	$\Delta D\mathrm{mean}$ ($\sigma$)	$\Delta D0.1\mathrm{cc}$ ($\sigma$)	$\Delta D99\text{\%}$ ($\sigma$)	$\Delta D\mathrm{mean}$ ($\sigma$)	$\Delta D0.1\mathrm{cc}$ ($\sigma$)
Brachial Plexus (L)	548.5 (55.4)	502.7 (20.2)	484.3 (45.8)	−418.5 (95.8)	−480.9 (43.6)	−415.5 (140.4)
Brachial Plexus (R)	547.3 (132.3)	489.8 (6.0)	532.0 (44.1)	−468.3 (151.3)	−444.1 (44.7)	−463.1 (88.4)
Brainstem	465.2 (159.8)	505.6 (7.2)	493.1 (23.4)	−238.6 (349.0)	−498.1 (8.8)	−440.7 (104.0)
Oral Cavity	528.5 (187.3)	490.9 (9.8)	305.1 (273.0)	−388.9 (278.2)	−507.8 (17.1)	−427.3 (197.5)
Cochlea (L)	496.2 (18.3)	459.1 (109.2)	455.3 (97.7)	−437.4 (91.3)	−448.2 (63.6)	−350.7 (101.5)
Cochlea (R)	460.5 (71.9)	494.9 (24.4)	486.6 (172.9)	−406.7 (119.0)	−448.6 (72.2)	−370.5 (154.7)
Esophagus	246.2 (249.1)	499.7 (31.1)	552.7 (141.8)	−143.8 (162.4)	−484.2 (28.0)	−528.9 (87.3)
Larynx	528.9 (96.5)	505.1 (9.9)	448.7 (76.2)	−667.6 (187.2)	−504.4 (27.2)	−449.3 (217.3)
Lens (L)	511.3 (214.0)	407.6 (85.0)	318.2 (46.5)	−84.8 (138.1)	−235.1 (346.2)	−33.8 (33.4)
Lens (R)	549.9 (107.8)	378.9 (7.9)	266.1 (5.8)	−79.3 (66.7)	−135.0 (113.8)	−116.0 (76.5)
Lungs	597.6 (320.2)	509.4 (26.6)	515.6 (152.8)	−363.8 (304.3)	−505.1 (20.4)	−487.0 (57.7)
Mandible	576.4 (26.4)	466.5 (26.2)	370.0 (6.6)	−653.2 (61.0)	−489.1 (17.3)	−287.5 (142.4)
Parotid (L)	577.1 (69.0)	502.6 (13.8)	344.4 (167.8)	−271.0 (183.0)	−484.7 (35.2)	−230.2 (207.0)
Parotid (R)	601.8 (79.4)	498.5 (10.7)	454.5 (194.9)	−384.2 (285.5)	−479.4 (32.9)	−417.2 (234.9)
Spinal Cord	501.4 (264.1)	517.3 (16.3)	529.0 (26.9)	−460.7 (361.2)	−495.6 (28.9)	−490.1 (65.4)
Submandibular Gland (L)	563.8 (75.4)	495.3 (14.6)	433.9 (91.8)	−348.5 (73.7)	−507.7 (0.9)	−608.6 (86.8)
Submandibular Gland (R)	531.2 (62.2)	503.9 (4.4)	468.1 (65.0)	−513.5 (94.0)	−502.6 (37.2)	−463.4 (188.9)

Fig. 2.

Example of generating a suboptimal dose from a clinical plan and then improving the plan quality. The clinical plan (upper left) is degraded with the prompt “Increase the mean dose to the esophagus by 800 cGy” (upper center), which is subsequently improved by the prompt “Decrease the mean dose to the esophagus by −800 cGy” (upper right) to be similar to the clinical plan. The DVH comparison (lower left) and dose difference maps show the overall extent of changes made and the return to clinic-like dose. The red arrow denotes the region of the esophagus.

Equivalence and error testing was performed for all 11 test-set patients, comparing each prompted change to minimum, mean, or maximum dose against TPS-calculated dose. The median RMSE was 52 cGy across all distributions. Predicted distributions achieved 98% equivalence with TPS dose under the RMSE criterion, and 99% under TOST p-values.

OOI-level aggregation (Table 3) showed general equivalence, with $\le$7 non-equivalent dose distributions per OOI. Additional dose-metric assessment showed modest (generally $\le$100 cGy) differences in OOI mean dose and D0.1cc (near-maximum). Target coverage was not compromised, with mean reduction of 8 cGy (D95%) and 1 cc (V100%).

Table 3.

Results of dose distribution equivalence testing, when prompting to make changes to the specified OOIs. Counts represent dose-level equivalence between dose distributions generated by the models to those calculated in a TPS using the RMSE equivalence criterion. The remaining four columns show mean differences between these dose distributions.

OOI (Left/Right)	Equivalent [count]	Not equivalent [count]	OOI $\Delta$Dmean [cGy]	OOI $\Delta$D0.1cc [cGy]	PTV $\Delta$D95% [cGy]	PTV $\Delta$V100% [cc]
Brachial Plexus (L)	58	2	107.4	−95.3	3.6	−1.2
Brachial Plexus (R)	73	5	73.3	−56.0	−1.4	−1.1
Brainstem	216	0	21.5	−130.0	−1.1	−0.4
Oral Cavity	209	7	−145.3	−155.5	−7.1	−1.2
Cochlea (L)	180	0	−5.7	0.5	−2.9	−1.1
Cochlea (R)	180	0	1.3	−17.5	−5.8	−1.6
Esophagus	254	1	−26.1	−192.1	−9.9	−1.1
Larynx	206	7	−118.6	−74.8	−5.3	−1.2
Lens (L)	53	7	−12.5	38.0	1.4	−0.6
Lens (R)	60	0	−0.8	12.6	−1.1	−0.7
Lungs	176	4	1.8	−168.6	−11.2	−1.5
Mandible	60	0	−65.9	−42.1	4.8	−0.0
Parotid (L)	239	4	−117.1	−60.7	−11.0	−0.9
Parotid (R)	209	4	−78.7	−108.8	−10.6	−1.1
Spinal Cord	207	3	−3.3	−95.3	−3.8	−0.6
Submandibular Gland (L)	36	6	−112.2	−270.2	−67.1	−1.9
Submandibular Gland (R)	107	1	−84.1	−177.4	−31.1	−1.1

A patient-level aggregate analysis addressed within-patient clustering. Across the 11 test patients, equivalence remained high, with only a small number of non-equivalent RMSE comparisons (Table 4). Patient-specific mean RMSE values ranged from 41.9–121.3 cGy. Visual inspection of randomly-selected dose distributions and DVHs confirmed similarity between predicted and TPS-generated doses.

Table 4.

Results of equivalence testing on a patient level of dose generated by the TPS to dose generated by the virtual dosimetrist model. Each count represents a dose-level equivalence result using the RMSE equivalence criterion.

	Equivalent [count]	Not equivalent [count]	Mean RMSE (95% CI) [cGy]
Patient 1	312	6	65.2 (59.3, 72.5)
Patient 2	135	0	61.2 (55.8, 66.9)
Patient 3	144	12	121.3 (104.2, 145.5)
Patient 4	390	9	59.1 (55.2, 65.2)
Patient 5	277	8	91.1 (84.4, 98.8)
Patient 6	263	1	48.2 (45.2, 51.6)
Patient 7	221	4	67.2 (61.4, 74.9)
Patient 8	144	3	59.8 (53.6, 68.5)
Patient 9	175	5	81.3 (70.5, 97.6)
Patient 10	266	1	59.9 (55.9, 64.6)
Patient 11	196	2	41.9 (37.7, 48.7)

4. Discussion

We introduced a text-conditioned deep learning framework to directly modify radiotherapy dose distributions of varying quality. Models understand prompts specifying absolute change to OOI minimum, mean, or maximum dose. Similar behavior under simple rephrasing suggests robustness, supported by CLIP’s broad language representation [11]. To our knowledge, this is the first deep learning-only, language-prompted dose modification system designed to bi-directionally adjust plan quality with potential for educational applications.

Despite training on templated input prompts, evaluation showed overall robustness to several unseen prompt variations. This possibly stems from using a language encoder pretrained on internet-scale data [11]. Alternative techniques like vectors specifying OOI name, metric, and change could have been used. However, vector inputs restrict formatting, whereas language encoders allow more natural model interactions.

CLIP was selected in part for its image–text design [11], conceptually similar to our volume-text approach, and strong medical task performance [12], [22]. The constrained language of short instructions describing dose metric, anatomical structure, and magnitude of change align with the text the encoder handles well.

Training prompt OOI names underwent simple standardization, such as “Left Parotid” instead of “Parotid_L”. Additionally, expanded OOI masks provided spatial context and controlled training loss (1). Inclusion of the OOI name in the prompt improved early-stage accuracy, suggesting the final models learned from both volumetric and language information.

Previous AI-based “virtual treatment planners” [23], [24], [25], [26] and LLM-based planning agents [27], [28] typically rely on a TPS for dose calculation. In contrast, our approach directly modifies dose distributions with deep learning models. This eliminates TPS dependency and reduces turnaround time to seconds.

A recently-proposed framework combines graph neural networks and LLMs for dose prediction and optional clinical interaction through natural language [29]. Language inputs serve as a refinement interface for high level post-generation adjustments. Similarly, Treat [30] uses language-guided protocol descriptors to condition dose generation across heterogeneous datasets. In contrast, our work uses natural language explicitly to modify dose distributions directly. This supports generating and rapidly refining suboptimal dose distributions.

The intentional generation of suboptimal dose distributions also addresses gaps in radiotherapy education. Clinic-based exposure often limits trainees to finalized plans without illustrating trade-offs or how planning decisions impact the final output. The iterative, language-guided approach developed in this work could provide a realistic, interactive environment for practicing plan review and edit skills.

Evaluations demonstrated reliable model responses to bidirectional prompts across diverse OOIs. Performance varied by structure and direction and was most reliable for mean dose changes (Table 2). Smaller changes for the oral cavity, esophagus, lenses, and mandible reflect anatomical complexity and learned constraints. For example, models appropriately resisted unrealistic changes to the esophagus, which typically lies outside high-dose regions. Similarly, a 500 cGy lens dose decrease often resulted in smaller reductions. This is reasonable, as the lenses are small and typically received low doses. Thus, a reduction of 500 cGy is not a typical planning objective, and the models resisted this unrealistic decrease.

Dose changes also extended to nearby structures, such that increasing dose to one OOI affected adjacent tissues, and unrealistic decreases appropriately compromised target conformity. These behaviors indicate the models internalized clinically plausible dose–response relationships and trade-offs, supporting their educational value.

Multicriteria optimization (MCO) similarly explores trade-off solutions. While educationally useful, MCO typically requires TPS integration and focuses on competing optimized plans rather than intentionally degraded ones. In contrast, our approach offers a single text-conditioned framework that rapidly generates suboptimal and improved distributions without TPS dependence.

4.1. Strengths and limitations

A primary strength is specifying dose change through natural language. This provides a single system to generate suboptimal examples (e.g., increasing OOI dose) and subsequently improve dose quality (e.g., reducing OOI dose). These techniques also generated TPS-comparable dose distributions without TPS integration, resulting in substantially faster dose updates, typically 7 s or less on appropriate hardware.

These strengths could enable future educational integration. The ability to arbitrarily change dose around OOIs could benefit educational simulators for plan-review practice. Similarly, the rapid response time could let trainees immediately visualize planning directive consequences. TPS independence improves global accessibility, including low- and middle-income countries. Clinically, this approach could generate alternative dose distributions for comparison or aid decision-making in adaptive workflows.

There are some limitations associated with this work. The current prompt design supports single-objective directives for minimum, mean, or maximum dose. Expanding to multi-objective prompts or additional metrics could increase applicability. While HN cases were chosen for their complexity, future work could evaluate other sites. Generating multiple modifications per patient meant individual plan-level equivalence results lacked statistical independent, a limitation addressed by the accompanying patient-level aggregate analysis. While full machine deliverability was not assessed, high equivalence to TPS-generated dose supports confidence in clinically realistic predictions.

All data originated from a single comprehensive cancer center. Although architecture principles should be robust, generalization to other institutional preferences, contouring protocols, dose grid resolutions, structural naming conventions, and vendors was not assessed.

Finally, the loop-based inference approach improved accuracy but introduced potential error accumulation or output instability. Identical inputs consistently yielded identical results. However, iteration-level convergence and error propagation were not independently evaluated.

In summary, this deep learning framework generates and iteratively refines dose distributions of varying quality in response to natural language. Modifications require seconds and are produced without TPS reoptimization or dose calculation at inference, offering a rapid, practical framework for radiotherapy plan review training. Real-time creation and refinement of suboptimal examples provides an approach that may support future applications in both educational and clinical environments.

CRediT authorship contribution statement

Skylar S. Gay: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Tucker J. Netherton: Writing – review & editing, Supervision, Formal analysis, Data curation, Conceptualization. Barbara Marquez: Writing – review & editing, Methodology, Conceptualization. Raymond Mumme: Writing – review & editing, Software, Conceptualization. Mary Gronberg: Writing – review & editing, Software, Conceptualization. Brent Parker: Writing – review & editing, Supervision, Formal analysis, Conceptualization. Chelsea Pinnix: Writing – review & editing, Supervision, Formal analysis, Conceptualization. Sanjay Shete: Writing – review & editing, Supervision, Formal analysis, Conceptualization. Carlos E. Cardenas: Writing – review & editing, Supervision, Formal analysis, Conceptualization. Laurence E. Court: Writing – review & editing, Supervision, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Conceptualization.

Ethics declaration

The authors have NOT obtained informed consent from participants or their legal representatives. Consent was waived due to the retrospective nature of the study.

This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. The University of Texas MD Anderson Cancer Center granted an exemption. The authors provided the following explanation: This study was exempt due to the retrospective nature of the study.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary material related to this article.

Data availability statement

The training dataset is not publicly available due to data size and institutional rules. A subset of the testing dataset is available from the corresponding author on reasonable request.