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

Simon Arberet, Riqiang Gao, Martin Kraus, Florin C. Ghesu, Wilko Verbakel, Mamadou Aliou Diallo, Anthony Magliari, Venkatesan Karuppusamy, Sushil Beriwal, Requite consortium, Ali Kamen, Dorin Comaniciu

TL;DR

AIRT (Artificial Intelligence-based Radiotherapy) is a new deep-learning framework that generates clinically deliverable VMAT radiation treatment plans in under one second—a process that typically takes minutes or hours. By using a differentiable dose-feedback loop and adversarial training, the system matches the quality of current gold-standard clinical software while enabling real-time, interactive plan adjustments.

What Problem Does This Paper Address?

Radiation therapy (RT) planning is a complex balancing act: clinicians must deliver a lethal dose of radiation to a tumor while sparing surrounding healthy organs (Organs-at-Risk, or OARs). Currently, this involves "inverse planning," where dosimetrists iteratively tweak optimization parameters in a Treatment Planning System (TPS). This process is time-consuming, leads to variability between planners, and creates a bottleneck in clinical workflows.

While previous AI attempts have tried to automate this, they often fall into two traps:

  1. Speed vs. Quality: Many models still require iterative "calls" to a heavy TPS, taking several minutes.
  2. Deliverability: AI often predicts "ideal" dose maps that the physical hardware (the Multi-Leaf Collimator or MLC) cannot actually execute.

Methods

The authors propose a modular, end-to-end pipeline that moves from raw CT images to a final machine-ready plan in a single forward pass. Key technical innovations include:

  • Differentiable Dose Feedback: Unlike "open-loop" models that predict a plan and stop, AIRT includes an internal, physics-informed deep learning dose engine. This allows the network to calculate the dose of its own predicted plan, identify errors, and apply a "correction" pass—all within the same sub-second inference window.
  • Adversarial Fluence Shaping: To ensure the plans are physically deliverable, the authors used a GAN (Generative Adversarial Network) approach. A "discriminator" was trained to distinguish between AI-generated plans and real clinical plans, forcing the generator to produce fluence maps with the specific patterns required for leaf sequencing.
  • User-Controlled Adaptability: The framework allows clinicians to use "sliders" at inference time to adjust the trade-off between target coverage and OAR sparing without needing to re-run an optimization loop.

Datasets

The model was trained and validated on a massive dataset of prostate cancer cases:

  • Primary Data: 1,277 CT scans (primarily from the REQUITE prostate dataset).
  • Data Augmentation: To improve robustness, the authors used a custom pipeline to expand the training set to 12,302 plans using randomized scaling, rotations, non-rigid deformations, and margin adjustments.
  • Evaluation: The model was tested against 62 validation cases and compared directly to RapidPlan Eclipse, a widely used clinical knowledge-based planning tool.

Key Insights and Ideas

  • The Power of "One-Shot" Correction: The researchers found that you don't need dozens of iterations to reach a high-quality plan. A single, differentiable feedback step is sufficient to bridge the gap between a "rough guess" and a clinically acceptable plan.
  • Bypassing the TPS: By embedding a simplified version of the physics equations (via the DL dose engine) directly into the neural network, the system eliminates the need to communicate with external, slow clinical software during the planning phase.
  • Real-Time Interaction: Because the inference is so fast (<1s), this technology could transform the clinician's role from "waiting for the computer" to "interactively exploring trade-offs" in real-time with the patient.

Key Results

  • Inference Speed: The entire pipeline runs in 889ms on a single NVIDIA A100 GPU. For context, traditional optimization typically takes 3 to 5 minutes.
  • Dosimetric Quality: AIRT demonstrated non-inferiority to RapidPlan Eclipse.
  • Target Homogeneity: The Homogeneity Index (HI) for AIRT plans was $0.10 \pm 0.01$, nearly identical to the clinical reference plans when evaluated using the industry-standard AcurosXB engine.
  • Clinical Validation: Independent testing by the American Oncology Institute (AOI) in India confirmed that the AI-generated plans passed standard patient-specific quality assurance (QA) for deliverability.

Limitations and Future Work

  • Anatomical Scope: The current model is specialized for "intact prostate" (prostate only, no lymph nodes). Adapting this to more complex sites like the Head & Neck or Lung will require handling more OARs and potentially multi-arc VMAT protocols.
  • Grid Resolution: To maintain sub-second speeds, the internal dose engine operates on a 4mm grid. While sufficient for many cases, clinical standards often prefer 2mm or 2.5mm resolutions.
  • Hardware Specificity: The model was trained for the Varian Millennium 120 MLC. Using it with different machines (like those with different leaf widths) would require retraining or a more flexible "foundation model" approach.
  • Simplified Physics: The internal dose engine uses a simplified source model. Future iterations could integrate more complex radiation transport physics to better align with high-end clinical solvers.

Summary updated

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.
BibTeX
@misc{arberet2026ai,
  author = {Simon Arberet and Riqiang Gao and Martin Kraus and Florin C. Ghesu and Wilko Verbakel and Mamadou Aliou Diallo and Anthony Magliari and Venkatesan Karuppusamy and Sushil Beriwal and REQUITE consortium and Ali Kamen and Dorin Comaniciu},
  title = {{AI End-to-End Radiation Treatment Planning Under One Second}},
  year = {2026},
  eprint = {2603.06338},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06338},
}
Figure