🔬 physics

One-for-All Adaptive Radiotherapy Planning Agent: A Foundation Framework for Daily CBCT-guided Radiotherapy

This paper introduces the "One-for-All Adaptive Radiotherapy Planning Agent," a unified foundation-model framework that autonomously generates clinically acceptable, treatment-specific online adaptive radiotherapy plans from daily CBCT scans in under two minutes across diverse cancer types while maintaining a human-in-the-loop oversight mechanism.

Original authors: Shaoyan Pan, Kirk Jon Luca, Yuan Gao, Shansong Wang, Mingzhe Hu, Ryan Sanford, Mojtaba Safari, Justin Roper, Zhen Tian, Tonghe Wang, Xiaofeng Yang

Published 2026-07-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Shaoyan Pan, Kirk Jon Luca, Yuan Gao, Shansong Wang, Mingzhe Hu, Ryan Sanford, Mojtaba Safari, Justin Roper, Zhen Tian, Tonghe Wang, Xiaofeng Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to hit a moving target with a slingshot. In the world of cancer treatment, that target is a tumor, and the slingshot is a beam of radiation. The goal is to blast the tumor with enough energy to destroy it while carefully avoiding the healthy organs nearby, like a surgeon trying to remove a splinter without nicking the skin. For decades, doctors have planned these attacks using a single "snapshot" of the patient's body taken before treatment begins. But here's the catch: bodies are not statues. Tumors shrink, organs shift, and people lose weight during the weeks of treatment. By the time the radiation is actually delivered, the "snapshot" might be wrong, meaning the beam could miss the target or hit something it shouldn't.

To fix this, doctors use a technique called "adaptive radiotherapy," which is like taking a new photo right before every single shot to adjust the aim. However, doing this manually is a nightmare. It requires a team of experts—doctors, physicists, and technicians—to spend hours every day re-drawing maps of the body and recalculating the beam's path. It's so slow and exhausting that most hospitals can't do it every day, leaving patients with plans that might be slightly off. This paper introduces a new kind of digital assistant, an "AI Agent," that acts like a super-fast, super-smart co-pilot. It can look at the daily photo, understand what has changed, and instantly redraw the entire battle plan in under two minutes, making daily adjustments possible for everyone.

The One-for-All Agent: Your Radiotherapy Co-Pilot

Meet the One-for-All Adaptive Radiotherapy Planning Agent (or A-RPA for short). Think of it as a highly skilled, digital Swiss Army knife designed specifically for cancer radiation treatment. In the past, if a doctor wanted to adjust a radiation plan because a patient's anatomy had changed, they had to run a series of separate, slow processes: first, they had to fix the blurry daily X-ray to look like a clear CT scan; then, they had to manually trace the tumor and organs; then, they had to calculate where the radiation would go; and finally, they had to check if the new plan was safe. This was like trying to build a house by asking one person to lay bricks, another to paint, and a third to install the roof, with each waiting for the previous person to finish.

This new Agent changes the game by doing it all at once. It is built on a "foundation model," which is like a super-intelligent brain that has already learned the rules of anatomy and radiation from thousands of examples. When a patient walks into the treatment room, the machine takes a quick 3D X-ray (called a CBCT) of their body. The Agent then asks the doctor, "Is the old plan still okay?"

If the doctor says, "Yes, the tumor hasn't moved much," the Agent quickly checks the old plan against the new X-ray and says, "All clear, let's go." But if the doctor says, "No, the tumor has shifted," the Agent doesn't just guess; it springs into action. It instantly creates a high-quality "synthetic" CT scan from the blurry X-ray, redraws the outlines of the tumor and sensitive organs, and predicts exactly where the radiation dose will land. It does all of this in under two minutes.

How the Magic Works (Without the Magic Wand)

The paper explains that this isn't just one big program doing everything blindly. It's more like a conductor leading an orchestra. The "Agent" part is the conductor, listening to the doctor's needs and deciding which tools to use.

The Synthesis Tool: If the daily X-ray is too fuzzy to see details, the Agent uses a generative AI (like a digital artist) to paint a clear, detailed CT scan based on the fuzzy X-ray.
The Registration Tool: This is like a smart map overlay. It takes the old plan (from before treatment started) and stretches or bends it to fit the patient's body shape today.
The Segmentation Tool: This is the auto-drawing feature. Instead of a human spending an hour tracing the tumor, the AI draws the new boundaries in seconds.
The Dose Tool: This predicts exactly how the radiation beam will scatter through the new body shape.

What makes this special is that all these tools are trained together. Usually, if you train a tool to make a pretty picture, it might not be good at measuring things. But because this Agent was trained to do all the steps together, the "pretty picture" it makes is actually perfect for the next step of measuring the dose. It's like a chef who not only chops the vegetables perfectly but also knows exactly how long to cook them, ensuring the final dish is delicious.

The Results: Fast, Accurate, and Safe

The researchers tested this Agent on a huge variety of patients, including those with cancers in the head, neck, lungs, abdomen, and prostate. They used both standard photon radiation (the most common type) and proton therapy (a more precise, high-tech type).

The results were impressive. When the Agent created a new plan, the radiation dose it predicted was incredibly close to the "gold standard" plans made by human experts.

For the tumor, the difference in the radiation dose was generally within 2.0 Gy (a unit of radiation dose) of the reference plan. In many cases, it was even closer, within 0.5 to 1.8 Gy.
The plans were so good that human experts reviewed them and found that 93% to 99% of the photon plans and 94% to 96% of the proton plans were clinically acceptable.
Even better, in about 34% to 50% of the photon cases and 28% to 54% of the proton cases, the Agent's plan was actually better than the original plan, offering improved protection for healthy organs or better coverage for the tumor.

The paper is careful to note that this isn't a magic wand that replaces doctors. The Agent still needs a human to look at the plan and give the final "thumbs up." It's a "human-in-the-loop" system, meaning the AI does the heavy lifting and the math, but the doctor makes the final call.

What It's Not (And What It Can't Do Yet)

The paper is very honest about its limits. The Agent works best when the daily X-ray is clear. If the X-ray is full of artifacts (like streaks from metal implants) or if the machine doesn't scan the whole body, the Agent might struggle. It also suggests that while the AI predicts the dose very well, it doesn't replace the complex physics calculations done by the hospital's main planning system; it acts as a fast, smart preview that gets the doctor 90% of the way there instantly.

Furthermore, the paper argues against the idea that you need a different AI for every single task. They tested this against other methods that tried to do just one thing (like just making the image clearer, or just drawing the lines). The "One-for-All" approach won because it learned how all the steps connect. A tool that just makes a pretty image might look good but fail to help the doctor calculate the dose accurately. By learning everything together, the Agent ensures that every step supports the next.

The Big Picture

This paper suggests that we are moving toward a future where radiation therapy can be truly "adaptive" every single day. Instead of a slow, manual process that takes hours and is only done occasionally, this Agent offers a way to adjust the treatment plan in the time it takes to brew a cup of coffee. It doesn't just speed things up; it makes the treatment more precise, ensuring that the radiation hits the tumor exactly where it needs to, every single time, while sparing the healthy tissue. It turns a complex, labor-intensive puzzle into a streamlined, intelligent conversation between the doctor and the machine.

Technical Summary: One-for-All Adaptive Radiotherapy Planning Agent

Problem Statement

Adaptive radiotherapy (ART) aims to deliver precise radiation doses by adjusting treatment plans to daily anatomical changes, such as tumor regression, organ deformation, or weight loss. However, current clinical practice relies on a single pre-treatment CT scan, leading to potential mismatches between the planned and delivered dose. While daily ART could mitigate this, it remains clinically impractical due to the resource-intensive, labor-heavy workflow. A typical daily adaptation requires complex, iterative decision-making involving new imaging, deformable registration, re-segmentation of targets and organs-at-risk (OARs), and dose re-optimization. This process often demands hours of expert input from radiation oncologists, medical physicists, and dosimetrists, creating a barrier to routine implementation. Existing AI solutions address isolated tasks (e.g., segmentation or dose prediction) but lack the higher-level clinical intelligence to orchestrate these components into a coherent, patient-specific adaptive workflow within the tight time constraints of treatment delivery.

Methodology

The authors propose the One-for-All Adaptive Radiotherapy Planning Agent (A-RPA), an end-to-end Agentic AI system designed to perform complete, treatment-specific online adaptive planning directly from daily cone-beam CT (CBCT) in under two minutes. The system comprises two primary components:

Unified Foundation Model (One-for-All):
- Architecture: Built upon a pretrained foundation model (DINOv2) for anatomical representation and a latent diffusion framework (Stable Diffusion 3.5) for generative tasks.
- Core Tasks: The model unifies four critical planning functions within a single framework:
  - Unsupervised CBCT-to-CT Synthesis: Generates synthetic CT (sCT) from daily CBCT using text-conditioned latent diffusion to ensure anatomical realism and dose-relevant features.
  - Multimodal Registration: Performs deformable alignment of prior planning data (CT, MRI, PET-CT) to the daily anatomy using a DINOv2-based encoder with 3D adapters.
  - Segmentation: Delineates tumor targets and OARs on registered or synthetic images, adapting prior contours to current anatomy.
  - Dose Prediction: Predicts 3D dose distributions conditioned on the synthesized anatomy, updated contours, and prior dose/metadata.
- Training Strategy: The model undergoes a two-stage training process. First, components are fine-tuned separately on large public datasets for specific tasks. Second, the entire framework is jointly post-trained on an internal clinical cohort (814 treatment courses) using a cascaded supervision strategy. This ensures that upstream representations (e.g., sCT) are optimized not just for visual fidelity but for downstream utility in segmentation and dose prediction.
Agentic Orchestration Layer:
- Controller: A Large Language Model (T5-XXL) acts as the "brain" of the system, interpreting clinician queries and managing the workflow.
- Decision Logic: Instead of a rigid pipeline, the agent dynamically selects the appropriate pathway based on the clinical scenario and observed anatomy. It can choose between:
  - Simple plan reassessment (propagating contours/dose).
  - Multimodal registration for improved visualization.
  - Re-segmentation on registered images.
  - Full replanning involving sCT synthesis, re-segmentation, and dose prediction.
- Human-in-the-Loop: The agent presents intermediate and final outputs for clinician review, allowing for intervention at critical decision points before final approval.

Key Contributions

Unified Framework: The first system to integrate synthesis, registration, segmentation, and dose prediction into a single foundation model, enabling end-to-end optimization where intermediate representations are tailored for downstream clinical tasks.
Agentic Workflow: A novel "human-in-the-loop" agent that autonomously determines the necessary adaptation pathway (from simple reuse to full replanning) based on daily anatomical changes, rather than executing a fixed sequence of tasks.
Joint Optimization: Demonstrates that jointly training synthesis and downstream tasks yields significantly better clinical utility (segmentation accuracy and dose prediction) compared to optimizing these modules in isolation or using competing unsupervised synthesis methods.

Results

The framework was evaluated on a diverse dataset spanning head-and-neck, lung, abdominal, and prostate cancers, covering both photon and proton therapy.

Dosimetric Accuracy: A-RPA generated adaptive plans with clinically acceptable accuracy. Target dose errors (D98) were generally within 2.0 Gy of reference plans. In photon therapy, PTV D95 and D98 errors ranged from 0.5–1.8 Gy and 0.7–2.0 Gy, respectively. Proton therapy showed similar stability.
Clinical Acceptability: Expert review classified 93–99% of photon cases and 94–96% of proton cases as clinically acceptable (Rank 2–5). A significant portion (34–50% for photon, 28–54% for proton) showed dosimetric improvements over the original plan (Rank 4–5).
Superiority of Joint Optimization:
- Synthesis: While competing unsupervised methods (SynDiff, ACSB) produced visually plausible sCTs, A-RPA's sCT resulted in significantly lower tumor dose errors (~1.2 Gy vs. ~8–12 Gy for competitors) and better OAR dose recovery.
- Segmentation: On synthetic CT, A-RPA maintained robust segmentation performance (Dice ~0.71–0.82), whereas competing methods degraded significantly. This led to a marked dose improvement of ~3.63 Gy for A-RPA compared to <1.0 Gy for competitors when re-segmentation was used.
Multimodal Registration: Incorporating MRI and PET-CT via deformable registration improved contour alignment (Dice increased to ~0.98 for body, ~0.95 for tumors) and downstream segmentation, particularly in soft-tissue structures where CBCT contrast is limited.

Significance and Claims

The paper claims that A-RPA represents a paradigm shift from isolated task-specific AI tools to a unified foundation-model approach for radiotherapy. Its primary significance lies in making daily online adaptive radiotherapy practically achievable by reducing the workflow time to under two minutes while maintaining clinical quality.

The authors emphasize that the agentic framework is crucial because adaptive radiotherapy is not a linear process; it requires clinical reasoning to determine if and how to adapt. By orchestrating branching workflows, A-RPA functions as a clinical decision support system rather than a mere automation pipeline. Furthermore, the study validates the hypothesis that joint optimization across the entire adaptive planning cascade (synthesis $\to$ registration $\to$ segmentation $\to$ dose) yields superior clinical outcomes compared to optimizing individual components in isolation. The system preserves clinical oversight through a "human-in-the-loop" design, ensuring that generated plans are reviewed and approved by clinicians before treatment delivery.

The authors acknowledge limitations, including dependence on CBCT image quality (performance degrades with severe artifacts or limited fields of view) and the need for broader multi-institutional validation. They also note that dose prediction remains an approximation of inverse optimization, though it provides a clinically coherent starting point for rapid adaptation.