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

AI summary

Overview

Research area: AI-driven radiotherapy treatment planning (radiation oncology, medical physics, deep learning for medical imaging, specifically volumetric modulated arc therapy / VMAT).

Technical level: Intermediate to Advanced — the concepts are explained accessibly, but the paper assumes familiarity with treatment planning terminology (fluence maps, leaf sequencing, DVH metrics, dose engines) and deep learning modules.

Scope: The paper presents AIRT, an end-to-end deep learning pipeline that turns CT images and structure contours directly into deliverable single-arc VMAT prostate plans in under one second, with dosimetric quality compared against RapidPlan Eclipse plans.

What This Paper Is About

Automated radiotherapy planning systems today typically rely on iterative optimization loops or on reinforcement learning that repeatedly queries a treatment planning system (TPS), which means plan generation takes minutes and depends on planner expertise. This paper asks whether a single end-to-end deep learning pipeline can skip the TPS entirely and produce a clinically deliverable VMAT plan — from imaging and contours all the way through leaf sequencing — in under one second.

Key Contributions

1. A fully TPS-free end-to-end pipeline (AIRT) that chains auto-contouring, a dose proposer, a Bev2Fluence network, a differentiable dose engine, a single-pass dose-error correction module, a second Bev2Fluence correction network, and a rule-based leaf sequencer, producing 180 fluence maps and an exportable DICOM RT Plan.
2. A differentiable dose feedback mechanism that computes a 3D dose error map from predicted dose (optionally including user-supplied scalar OAR penalties), projects it into the beam's-eye view, and corrects the fluence maps in a single feed-forward pass rather than through multiple TPS refinements.
3. Adversarial fluence map shaping plus an ad-hoc VMAT leaf sequencer, used to keep predicted fluence maps on the manifold of deliverable, leaf-sequencable VMAT patterns.
4. Sub-second single-arc VMAT prostate planning with statistical non-inferiority to RapidPlan Eclipse demonstrated on target coverage and OAR sparing metrics, plus an ablation study isolating data augmentation, dose feedback, and the adversarial loss.

Main Findings

• Speed: The pipeline generates single-arc VMAT prostate plans, including leaf sequencing, in under one second on a single Nvidia A100 GPU. The paper reports this is faster by multiple orders of magnitude than RL + TPS pipelines requiring approximately 80 to 100 seconds including TPS refinement.
• Training data: The model was trained on more than 10,000 intact prostate cases. In the Methods, 1,277 CT patient scans were used (primarily the REQUITE prostate dataset of 1,001 scans plus smaller in-house datasets), split into 1,122 training, 62 validation, and 60 testing cases; data augmentation expanded the 1,122 training plans to 12,302.
• Prescription and delivery: Plans were created for intact prostate (PTV, CTV, and PTV ∩ Rectum all set to 40 Gy), delivered as a single arc VMAT with a Varian Millennium 120 (M120) MLC.
• Homogeneity with the clinical dose engine: Evaluated with Eclipse AcurosXB, target homogeneity was HI = 0.10 (± 0.01) for AIRT versus 0.10 (± 0.01) for Eclipse — described as similar. With the DL dose engine, AIRT had a lower mean PTV HI (0.11) than Eclipse (0.16), a difference the authors attribute to differences between the DL dose and AcurosXB rather than to an actual plan-quality improvement.
• Target coverage: With AcurosXB, PTV D98 was 39.3 Gy (± 0.2) for both Eclipse and AIRT. With the DL dose engine, PTV D98 was 38.4 Gy (± 0.8) for Eclipse and 38.8 Gy (± 0.4) for AIRT.
• OAR sparing: Organs-at-risk received comparable dose metrics with both methods. With the DL dose engine, rectum doses were slightly lower for AIRT while bladder doses were slightly higher; similar trends appeared for Rectum D50 under AcurosXB. For example, under AcurosXB, bladder D2 was 35.8 Gy (± 6.0) for Eclipse versus 37.0 Gy (± 5.9) for AIRT, and rectum D2 was 31.4 Gy (± 5.6) versus 32.5 Gy (± 6.0).
• Statistical non-inferiority: Tested on the 62 validation cases using Eclipse AcurosXB, with a margin of 0.01 for the homogeneity index and a margin of 1.5 Gy for all other dose metrics; AIRT met the non-inferiority margins at p < 0.05. Per-metric details are reported in Supplementary Table S3.
• Representative cases: DVH curves for six cases were selected by ranking the validation set by PTV volume and picking the 0th (minimum), 20th, 40th, 50th (median), 80th, and 100th (maximum) percentiles, to avoid cherry picking. AIRT maintained PTV coverage and OAR sparing across the full range of PTV sizes.
• Ablation study: Across variants evaluated on 62 validation cases with the DL dose engine, data augmentation systematically improved results. Removing GAN loss worsened PTV homogeneity (HI 0.148 for AIRT without GAN versus 0.107 for AIRT full), which the authors say indicates leaf-sequencability problems without adversarial loss. Dose feedback improved PTV homogeneity, but without adversarial loss it could push fluence maps off the deliverable manifold. Best results came from combining dose feedback and the adversarial loss. The authors note relative metrics can worsen when dose feedback is used because it targets absolute metrics.
• OAR sparing adaptability: The dose feedback mechanism accepts scalar penalties per OAR at inference time. Increasing the bladder or rectum sparing factor decreased dose in the corresponding organ with limited effect on the other, but decreased PTV homogeneity — the classical target-versus-OAR trade-off. In the figure legend, s_r = s_b = 0 is the baseline, and s_b = 2% means the dose feedback mechanism tried to decrease bladder dose by 2% voxel-wise compared to its input dose distribution.
• Independent clinical testing (unpublished): The American Oncology Institute (AOI) in India independently tested AI-generated plans, achieving target coverage comparable to manually crafted plans and passing patient-specific quality assurance testing for deliverability. The paper states these results are currently unpublished.

Methodology in Plain English

The pipeline takes a CT volume and contours (body, PTV, OARs) as input and produces a leaf-sequenced plan as output, one module after another:

1. Dose proposer — a 3D ResUNet predicts a clinically plausible 3D dose distribution from the CT and contours.
2. BEV projection — a parameter-free step projects that 3D dose into the beam's-eye view of each control point, aligning it geometrically with the fluence maps.
3. Bev2Fluence — a 3D convolutional network (MedNeXT backbone) predicts all 180 fluence maps jointly, since control points are strongly coupled by VMAT delivery constraints and because dose is the accumulation of every control point's contribution.
4. Differentiable dose engine — a physics-informed deep learning network computes dose from the predicted fluence maps so correction can happen at inference time. It was trained on an LTBE solver using transport physics similar to AcurosXB but with a simplified beam model, and it runs at 4 mm dose resolution to keep inference under one second. It is kept frozen.
5. Dose error module — a parameter-free module builds a 3D dose error map comparing predicted and desired dose, primarily penalizing PTV hot spots more than cold spots. It can optionally accept scalar inputs to penalize excess dose in each OAR, letting a clinician trade target homogeneity against OAR sparing.
6. Bev2Fluence correction — a second network refines the fluence maps using the initial prediction plus the BEV-projected dose error, closing the dose feedback loop in a single pass (one iteration, not an unrolled loop).
7. Leaf sequencing — a rule-based, Cython/C-implemented algorithm converts each control point's fluence map into left/right leaf positions and monitor units, modeling partial-pixel effects at moving leaf boundaries and the dosimetric leaf gap. It runs in parallel per control point.
8. RT Plan export — leaf sequences and MU values are exported as a DICOM RT Plan for review or delivery.

Training happens in two stages. Stage 1 trains the full pipeline with reconstruction L1 losses (dose proposer loss, dose loss, fluence map losses, dose error losses) and no adversarial loss, to establish stability. Stage 2 introduces an adversarial loss with a discriminator that distinguishes pipeline-generated fluence maps from Eclipse-optimized target fluence maps, with randomly generated user-control parameters passed to the dose error modules; the network before the correction module is frozen here.

Evaluation used four dose engines: the physics-informed DL engine, an in-house LTBE solver derived from the AcurosXB codebase, Eclipse AAA, and Eclipse AcurosXB. Only the DL dose engine and Eclipse AcurosXB results appear in the main manuscript; the other two are in the supplement. The comparison baseline was RapidPlan plans optimized in Eclipse with the Photon Optimizer and AcurosXB.

Why This Matters

Impact on research: The paper argues that end-to-end differentiability — with dose objectives optimized via backpropagation through the whole pipeline — offers an alternative to reinforcement learning and model predictive control approaches that depend on repeated TPS calls and episodic rewards. It demonstrates that a single feed-forward network, given a differentiable dose error correction module, enough network capacity, and a large training set, can produce deliverable fluence maps without an optimization loop. It also positions the dose feedback mechanism as an alternative to traditional Multi-Criteria Optimization, where trade-offs can be explored nearly instantaneously without replanning.

Real-world applications:

• High-volume centers could increase throughput by generating standardized, clinically deliverable plans almost instantly instead of running iterative TPS optimization.
• Regions with scarce qualified dosimetrists and physicists could gain access to advanced radiotherapy techniques that currently require planner expertise.
• Clinicians could run an interactive session, reviewing multiple candidate plans in real time and converging on an acceptable plan within a single appointment.
• Because the pipeline outputs DICOM RT Plans, it can be integrated into clinical systems, or the AI plan can be used as a warm-start for further manual refinement on special cases.

Industry relevance: The work comes from Siemens Healthineers researchers across Digital Technology and Innovation and Varian Medical Affairs, uses Varian Eclipse and the M120 MLC, and is positioned around integrating AI planning into existing clinical TPS workflows and DICOM-based delivery. The paper carries a disclaimer that the concepts are based on research results not commercially available, with no guarantee of future commercial availability.

Future Directions

• Broadening anatomical and technical scope: Extending beyond single-arc VMAT intact prostate to multi-arc VMAT, simultaneous integrated boost (SIB), and other body regions such as lung, breast, and head and neck — which would require adapting network inputs/outputs, generating large-scale datasets, and possibly multiple dose-feedback loops to handle the increased underdeterminacy.
• MLC generalization: The current model was trained and validated for the Varian Millennium 120 MLC; retraining may be needed for other MLC models, especially with different leaf widths, and a flexible foundation model addressing multiple configurations is suggested as a possibility.
• Better dose modeling: The DL dose engine uses a simplified source model with a spatially constant photon spectrum and no electron contamination, unlike AcurosXB, which uses a spatially varying spectrum and models secondary electrons. The authors say closer alignment of source and beam models could improve plan quality.
• Higher dose resolution: The 4 mm grid is coarser than typical TPS dose engines; progress in differentiable dose engine efficiency could enable higher resolution and potentially better PTV–OAR sparing or more targeted dose objectives like SIB.
• Multi-institutional validation: The authors state it is essential to validate performance across clinical institutions and anatomies; the unpublished AOI India testing is described as a step in that direction, with conference abstract and journal submission planned.

Target Audience

Medical physicists, radiation oncologists, dosimetrists, and clinical workflow researchers interested in automated treatment planning; machine learning researchers working on end-to-end differentiable pipelines, physics-informed neural networks, and generative or adversarial methods for medical physics; and industry or health-system stakeholders evaluating AI planning tools for clinical deployment and access-to-care problems.

Authors’ abstract

Artificial intelligence-based radiation therapy (RT) planning has the potential to reduce planning time and inter-planner variability, improving efficiency and consistency in clinical workflows. Most existing automated approaches rely on multiple dose evaluations and corrections, resulting in plan generation times of several minutes. We introduce AIRT (Artificial Intelligence-based Radiotherapy), an end-to-end deep-learning framework that directly infers deliverable treatment plans from CT images and structure contours. AIRT generates single-arc VMAT prostate plans, from imaging and anatomical inputs to leaf sequencing, in under one second on a single Nvidia A100 GPU. The framework includes a differentiable dose feedback, an adversarial fluence map shaping, and a plan generation augmentation to improve plan quality and robustness. The model was trained on more than 10,000 intact prostate cases. Non-inferiority to RapidPlan Eclipse was demonstrated across target coverage and OAR sparing metrics. Target homogeneity (HI = 0.10 $\pm$ 0.01) and OAR sparing were similar to reference plans when evaluated using AcurosXB. These results represent a significant step toward ultra-fast standardized RT planning and a streamlined clinical workflow.