List of Papers Browse by Subject Areas Author List
Abstract
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.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3613_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{PolIsa_Learning_MICCAI2026,
author = { Poles, Isabella AND Arberet, Simon AND Gao, Riqiang AND Kraus, Martin AND Santambrogio, Marco D. AND Ghesu, Florin C. AND Kamen, Ali AND Comaniciu, Dorin},
title = { { Learning to Optimize Radiotherapy Plans via Fluence Maps Diffusion Model Generation and LSTM-Based Optimization } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents an end-to-end Learning-to-Optimize (L2O) framework for Volumetric Modulated Arc Therapy (VMAT) planning. The framework integrates two novel components: a one-shot Fluence Map Diffusion (FMD) model that learns a clinically deliverable manifold, and an LSTM-based L2PLAN module that learns task-adaptive update dynamics. This architecture shifts the planning paradigm from slow, conventional iterative optimization to an efficient, diffusion-based inference process.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(1). The paper introduces a novel combination of distilled one-shot diffusion models and recurrent meta-optimizers to navigate the non-convex landscape of radiotherapy planning. (2). The LSTM-based L2PLAN module captures gradient history to learn update dynamics, allowing the optimizer to navigate complex dose landscapes more effectively than traditional fixed-rule optimizers. Experimental results demonstrate that L2PLAN achieves comparable performance to standard planners but with 7x to 20x fewer optimization steps. (3). The system provides a seamless transition from patient CT images to deliverable machine plans without manual intervention, significantly balancing planning speed with clinical flexibility.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(1). As the student generator G_θ is distilled from a teacher, it may inherit approximation errors. The paper does not quantify if the L2PLAN refinement can fully overcome these manifold errors or if the final plan quality is ultimately capped by the teacher’s fidelity. (2). The meta-optimizer is trained using k_i inner-loop steps. The manuscript provides no sensitivity analysis regarding the choice of k_i. In L2O research, the length of this unrolling horizon is a critical determinant of whether the optimizer learns “short-sighted” updates or achieves long-term stability. (3). The authors use a two-layer LSTM to track gradient history. In current L2O research, there is a shift toward Transformer-based architectures for long-range dependency tracking. The authors should clarify if the LSTM was chosen strictly for its lower computational footprint or if they observed performance plateaus with other architectures. (4). The experimental data is restricted exclusively to single-arc prostate plans with a fixed 30^° collimator and 0^° couch angle. The lack of evaluation across dual-arc/multi-arc configurations, hardware from other manufacturers, or different tumor sites like lung or head and neck suggests the claim of a “universal” method lacks sufficient experimental support. (5). While the REQUITE dataset is sizable, the descriptions of the two private prostate cohorts are brief. The paper lacks detail on the diversity of image protocols or specific preprocessing steps used to normalize CT scanner protocols across different institutions, which is vital for verifying institutional generalizability. (6). The regularization weights are fixed throughout the experiments without sensitivity analysis. Given that these weights balance dose fidelity against plan smoothness, the paper should explain if these were manually adjusted or if they require patient-specific tuning for different clinical objectives. (7). The methodology mentions using a “rule-based LS tool” and a “differentiable dose engine” but provides limited information on the specific software versions or proprietary platforms used for the clinical simulations. Specifying the platform used for the VMAT simulation would significantly enhance the transparency of the clinical translation evaluation. (8). The methodology section is rich in mathematical formulation. To improve accessibility for a broader, cross-disciplinary audience, it would be beneficial to bridge these technical details more explicitly with high-level conceptual summaries throughout the section.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper innovatively applies generative and meta-learning techniques to VMAT planning, achieving significant and clinically relevant improvements in speed and deliverability. However, it is recommended as a Weak Accept because the authors have not provided a sensitivity analysis for the unrolling horizon k_i, nor have they quantified the propagation of approximation errors from the teacher model or justified the use of a two-layer LSTM over modern Transformer-based architectures; additionally, the narrow clinical validation on a single tumor site and machine geometry lacks the experimental breadth necessary to support the claim of a universal VMAT planning solution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
In this paper, the authors are trying to solve a well-defined and practically important problem in VMAT planning. The design of the diffusion-based learning-to-optimize framework is pretty decent. By applying the teacher and student model structure, the authors distill a complex diffusion model into a one-shot generator. By using an LSTM-based meta optimizer, the proposed framework can learn adaptive gradient update dynamics to efficiently refine fluence maps toward prescribed dose objectives without requiring tuned initialization or retraining. Overall this innovative framework works well on different test data.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) Novel and clinically motivated problem formulation. 2) The LSTM-based meta-optimizer addresses a real limitation of existing approaches. Rather than relying on fixed-hyperparameter optimizers that require careful iterative tuning, the learned optimizer adapts its momentum parameters dynamically based on gradient history. 3) The experiments are comprehensive and well-designed. The authors isolate different parts to clearly show the actual gain and loss. Statistical significances are also reported in the paper. 4) The application of the teacher student model to distill heavy diffusion model.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) Evaluation limited to only one type of cancer treatment. 2) Miss some standard clinical dose evaluation metrics.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1) The framework contains three separate parts and a frozen DL dose engine, which is complex. Including a failure modes discussion would strengthen the paper’s depth and completeness. Let’s say if the FMD generator produces a fluence map at the boundary of the deliverable manifold, can LSTM optimizer hold such a condition correct? 2) The dataset is limited to prostate cancer which is relative simple planning scenarios with few OARs and limited geometric complexity. It is good to test model performance on a more complex scenario like head-and-neck cancer. 3) The authors report many image quality metrics like SSIM, PSNR, etc. in the paper. While these metrics are good signals to evaluate generative modeling, they do not show any clinical relevance dose distribution evaluations. It is good to add some DVH-based metrics like D95, D2, etc.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The authors propose diffusion-driven Learning-to-Optimize method for end-to-end VMAT planning. This method first uses a diffusion model to synthesize a clinically plausible fluence map, and then introduces an LSTM-based gradient optimizer to enable personalized optimization and adaptation of the fluence maps generated in the first stage. By iteratively updating the fluence maps, the method achieves clinical goals consistent with dose delivery constraints and improves planning efficiency.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The authors adopt the teacher diffusion model to obtain the fluence map in one-shot sampling by directly estimating the clean fluence map from noisy ones, which avoids the high computational cost brought by the iterative denoising process. 2.The authors propose a LSTM-based Learning-to-Optimize VMAT Plans module that learns iterative update dynamics to efficiently refine fluence maps toward prescribed dose objectives without tuned initializations or retraining. 3.Experiments on three prostate cancer clinical and public patient cohorts demonstrate the proposed method improves efficiency, flexibility to changing dose objectives, and plan deliverability.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The Fluence Map Diffusion model appears to be trained only as an unconditional diffusion model in the image domain. How does this ensure consistency between the synthesized fluence maps and the physical parameters of the machine? 2.The synthesis of fluence maps is performed via the one-shot sampling from a normal distribution, which may introduce significant uncertainty in the fluence maps, causing the second-stage optimization to start from a largely biased initial point. How do the authors address this uncertainty? 3.Some reported results seem not reliable enough. For instance, the authors use MAE to evaluate dose coverage in the PTV, which I consider a coarse metric because it is easily affected by extreme values. I recommend that the authors report clinically relevant metrics, such as D99.4.Some experimental details are missing, such as the number of sampling steps used during training of the Fluence Map Diffusion model. 5.There are some formatting oversights. For example, Table 2 contains a large number of blank areas. 6.There is some unclear use of italics in Section 3.5, as well as underlining without clear meaning on page 2.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed diffusion-driven learning-to-optimize method for VMAT planning is novel and shows clear advantages in efficiency, flexibility, and deliverability. However, several concerns limit a stronger recommendation: the unconditional diffusion training may introduce considerable uncertainty; evaluation relies on coarse quality metrics (e.g., MAE) rather than clinical ones.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank AC and Reviewers for their constructive comments. Reviewers recognize our method as “novel” (R1,R2,R4), “clinically motivated”, with “well-designed experiments” (R2) achieving gains in “efficiency, flexibility, and deliverability” (R4). However, they require clarifications about (A) generator fidelity/uncertainty and sampling (R1,R2,R4), (B) L2Plan and LS design and parameters (R1), (C) organs/machine generalization (R1,R2), (D) DVH metrics (R2,R4).
(A) We thank the Reviewers for raising this point. FMD mitigates errors starting from distillation by combining L_DMD and L_adv anchoring generated samples to real ones. Table 2 shows the 18-step teacher reaches 15.40 FID, while 1-step FMD with both losses reaches 17.52 FID. This residual influences final plan fidelity but is small enough to be absorbed by L2Plan rather than act as a ceiling; a regime weaker generators fall outside of. In addition, because the LSTM trains on samples from N(0,1) spanning both tail and error cases, it learns to navigate complex landscapes: correcting trajectories from atypical initializations and reducing convergence steps for in-distribution ones. The pre-/post-LS statistical agreement (Table 1, Fig. 2(a)) shows that even possible atypical generations across the test set remain deliverable. Future work will complement this analysis by only sampling error- and boundary-like cases for a more precise quantification under extreme starting conditions. Regarding machine consistency, FMD is unconditional but not unconstrained: it is trained on F_del, the manifold of machine-constrained fluence maps defined by leaf-motion, dose-rate, and gantry-rotation limits.
(B) We thank R1 for the constructive comments. k={5,10,20} analysis has been experimented: k=5 was short-sighted, k=20 added memory/time without gains, while k=10 is the stability–efficiency optimum, matching L2O findings on unroll length. We acknowledge that Transformers are promising. However, we chose a 2-layer LSTM for its lightweight structure and per-step latency, critical for efficient planning. Regularization weights were tuned on the validation set and fixed across all test cases (no patient-specific tuning). We also clarify implementation details: LS model is C-based parallel rule-based converting each CP fluence map into MLC apertures and MU with sub-pixel and dosimetric-leaf-gap corrections; the dose engine is a frozen physics-informed DL model trained on an LTBE solver, with spherical-harmonic fluence accumulation and a MedNeXT backbone at 4 mm.
(C) We agree validation consists of 3 prostate cohorts single-arc VMAT with fixed geometry and Varian M120 MLC and will revise to avoid overclaims. Future work will validate our model generalizability to multi-/dual-arc plans and complex sites. Preprocessing across all cohorts was limited to organ/OAR/helper-structure segmentation via AI-Rad Companion and an in-house C++ pipeline; plans were then created in Eclipse with the RapidPlan prostate SBRT model on single-arc VMAT with Varian M120 MLC. Training REQUITE CTs and contours were augmented by scaling, small rotations, CPAB non-rigid deformation, and clinically plausible PTV/rectum margin changes, expanding training set to 12,153 plans.
(D) We agree that DVH metrics are more clinically interpretable. We will specify them for the Fig. 2(c) cases: trade-off PTV [D95=35.79, D2=40.05, HI=0.16], Rectum [Dmean=4.59], Bladder [Dmean=8.84]; PTV-homogeneity PTV [D95=38.23, D2=39.52, HI=0.03], Rectum [Dmean=8.23], Bladder [Dmean=13.04]; OAR-sparing PTV [D95=30.46, D2=40.26, HI=0.22], Rectum [Dmean=3.60], Bladder [Dmean=4.52]. Full-cohorts DVH would require restructuring Table 1 beyond page limits; representative values will be added to the main text and full analysis to an extended version.
We hope we can clarify the concerns and provide a revised manuscript that solves “a real limitation of existing approaches” (R2) while “achieving significant and clinically relevant improvements” (R1).
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
All reviewers agree on the novelty of the approach with a clinically motivated problem formulation. Authors are invited to adresses the issues pointed out in the reviewers reports.