cs.AIOct 2, 2026

RadOnc-Agent: An LLM-Orchestrated Framework for AI Workflows Across the Radiotherapy Care Pathway

Authors: Caiwen Jiang, Shuoyang Wei, Songlin Zhao, Junyu Li, Jingyuan Chen, Wei Liu

Organizations: Department of Radiation Oncology, Mayo Clinic Arizona, Phoenix, AZ, USA.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Artificial intelligence has advanced individual radiotherapy tasks, yet these capabilities remain separated across clinical stages, software environments and data modalities. This fragmentation contrasts with the longitudinal radiotherapy workflow from treatment decision-making through follow-up. Here we present RadOnc-Agent, an agentic artificial-intelligence framework that formalizes radiotherapy into four clinical phases and provides 26 callable functions through a conversational interface. A large-language-model controller maps clinical intent to schema-constrained calls, preserves patient and workflow context, and routes requests to specialist services. We evaluated system execution using 2,600 single-function requests (7,800 repeat executions), 200 prespecified synthetic cross-stage scenarios spanning four phases (600 executions), and 120 workflow instances from 60 de-identified patient records (360 clean executions) representing decision-to-planning and planning-to-adaptation. RadOnc-Agent selected the intended function in 98.79% of single-function executions, completed 96.50% of scripted cross-stage workflows, and completed 96.67% of real-patient workflow executions. In comparative ablations, removing longitudinal state reduced cross-stage completion from 96.50% to 84.00%, while disabling schema and identity validation increased mismatched backend dispatch from 0% to 95.28% in a replay/test evaluation. These findings establish the technical feasibility of an LLM-orchestrated architecture for coordinating heterogeneous radiotherapy capabilities and information across longitudinal workflows; they do not establish clinical correctness, clinical utility or prospective benefit.

Figures & tables

Function groupCallable functionFunction descriptionHeadline validation resultEvidence status
Diagnosis & DecisionRT-indication assessmentRetrieves guideline-supported radiotherapy indications and fractionation recommendations with source citations.Top-5 retrieval 97.50%; overall answer accuracy 95.33%; unsupported-query withholding 96.67%.Current task-level evaluation
 Modality selectionCompares candidate radiotherapy modalities using patient context and, when available, verified quantitative risk parameters.Completely acceptable regimen 46–52%; acceptable including minor deviations 64–66%; modality-selection accuracy 70.5–72.2%.Previously published task-level evaluation
 
Simulation & PlanningPlanning directiveConverts diagnosis, treatment intent and prescription into a structured planning directive.Supporting order-summarization evaluation: more than 98% agreement with therapist-verified references.Previously published task-level evaluation
 MR-to-CT synthesisGenerates a planning-compatible synthetic CT from planning MRI.MAE 46.8 ±\pm 12.7 HU; SSIM 0.961 ±\pm 0.018.Completed internal task-level evaluation
 Deformable registrationRegisters longitudinal CT images and returns a deformation field and propagated anatomy.OAR/target Dice 92.07%/87.14%; OAR folding 0.08%; accumulated-dose MAE 0.72%.Completed internal task-level evaluation
Table S1: Functions and headline validation results for RadOnc-Agent. The table describes all 26 callable functions and reports concise, function-appropriate performance results where a task-level evaluation was required. Direct retrieval, external-API and static-response utilities are identified as having no separate predictive-performance endpoint. Dataset composition, development partitions and reference standards are reported in Supplementary Table S2 . Evidence status identifies whether the reported evidence was generated in the current study, reported previously, completed internally but not previously published, or limited to the current system-level technical evaluation; it does not denote clinical deployment or regulatory status.
Callable functionDevelopment or training dataEvaluation dataReference standard and endpoints
RT-indication assessmentBM25 retrieval plus general-language-model synthesis; no task-specific parameter training.300 prespecified questions across ten disease sites: 240 supported and 60 unsupported or out of corpus.Radiation-oncologist-adjudicated answers, supporting passages and refusal status; retrieval, answer, citation and withholding endpoints.
Modality selectionFine-tuned LLaMA2 and Mistral models; 7,500 curated development cases.Independent test set of 403 patient cases [ 11 ] .Radiation-oncologist assessment of regimen acceptability and agreement with the documented treatment modality.
Planning directiveGeneral-language-model order summarization; no task-specific parameter training.607 CT-simulation orders from seven treatment-modality and disease-site groups [ 12 ] .Therapist-verified order summaries; summary accuracy and consistency.
MR-to-CT synthesis2.5D U-Net++; public SynthRAD2025 release, with 462 training and 58 validation cases.58 patient-level held-out SynthRAD2025 MRI-to-CT cases [ 22 ] .Reference CT; regional HU MAE and SSIM.
Deformable registrationPatient-adaptive registration network; 2,052 patients/7,429 image pairs for training and 293 patients/1,062 pairs for validation.Internal test: 586 patients/2,122 image pairs; external test: 542 patients/1,373 pairs from two private and five public cohorts [ 13 ] .Expert contours and reference dose; Dice, folding and accumulated-dose MAE.
Auto-contouringTreatment-plan-conditioned three-dimensional segmentation network; 3,355 training and 387 validation plans.Pooled two-centre held-out test set of 370 plans from 279 patients, plus fully held-out external cohorts [ 14 ] .Physician-defined target and OAR contours; Dice and inference time.
Table S2: Development and evaluation data for functions with task-level validation. The table summarizes development or training data, evaluation data and reference standards. Model descriptions are limited to their primary implementation class. Direct record-retrieval, external-API and static-response utilities without a separate predictive endpoint are described in Supplementary Table S1 and are not repeated here.
Execution riskDetection mechanismFail-safe behaviorControl mode
Missing required inputFunction-specific schema checks identify absent required arguments before backend routing.Backend dispatch is blocked; the agent issues a clarification or withholding response and can resume the workflow only after the required information is available.Deterministic gate with LLM-mediated clarification
Unregistered function, malformed argument or invalid routeThe function registry validates the function name, required and optional arguments, and registered backend route.The call is rejected before backend invocation rather than routed to an unregistered or incompatible service.Deterministic
Patient, course or plan identifier mismatchIdentifiers in each call are checked against the active session; structured outputs retain patient-, course- and plan-level provenance.The inconsistent call is rejected and is not merged into the active patient/workflow context.Deterministic
Cross-patient context carryoverSession state is bound to one active patient and treatment course, with no persistent conversational memory or cross-patient cache; state is cleared between independent executions.Information from a different session is not retained as active patient or workflow context for subsequent calls.Deterministic session isolation
Transient backend or asynchronous-job failureBackend status, transport errors and asynchronous-job progress are monitored by the dispatch layer.Workflow state is retained for bounded retry or resumption after service restoration; an unresolved failure is returned explicitly rather than treated as a successful result.Hybrid deterministic handling and LLM-mediated resumption
Table S3: Workflow-execution safety controls and fail-safe behavior. The controls govern function calling, backend routing and session isolation; they do not establish the clinical correctness or clinical appropriateness of specialist-function outputs.
More in this topic

Workflow Automation

Workflow automation coordinates multi-step tasks across applications, tools, and people using software, AI agents, and explicit process logic.

See all 62 papers

Explore similar work

CardsList
Dec 12, 2024

How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?

Radiology, with its heterogeneous modalities, anatomies, and evolving protocols, is a natural yet high-stakes testbed for agentic AI. As LLMs grow more capable and tool ecosystems such as MCP and Agent Skills more complex, their capability boundaries in radiology remain unclear. We introduce RadA-BenchPlat, a two-layer benchmark spanning idealized reasoning and real-world execution: a synthetic layer with 2.2k clinician-verified records, 24.2k QA pairs, and 10 tool categories under diverse availability settings; and a real-environment layer pairing 165 2D/3D cases with executable tools for grounding, diagnosis, and report generation. We find that high completion rates under idealized settings do not carry over to real environments, where the best agents trail their upper bound by approximately 15% and drop to 0.327 on long-chain tasks. Prompting strategies (few-shot exemplars, multi-agent coordination) and AutoTB (on-the-fly synthesis of missing tools) mitigate execution drift and breakdowns in difficult scenarios, lifting execution success to 94.5%, yet remain not fully reliable. Overall, our findings suggest that radiology agents are no longer limited to conceptual demonstrations, but are beginning to show potential for real-world applications. Project repository: https://github.com/MAGIC-AI4Med/RadABench

Qiaoyu Zheng, Chaoyi Wu, Weike Zhao +5

Shanghai Jiao Tong University, Shanghai, China · Department of Radiology, Renmin Hospital of Wuhan University, Wuhan, China

HealthcareLLM Agent EvaluationAI Agent BenchmarksLLM AgentsLong-Horizon LLM Agents
AskPDF
Feb 17, 2026

Multi-Modal Multi-Agent Reinforcement Learning for Radiology Report Generation

We propose MARL-Rad, a multi-modal multi-agent reinforcement learning framework for radiology report generation that trains the entire agentic system on policy within its deployed radiology workflow. MARL-Rad addresses the limitation of post-hoc agentization, where fixed LLMs are organized into hand-designed agentic workflows without being optimized for their assigned roles. Our framework decomposes chest X-ray interpretation into region-specific agents and a global integrating agent, and jointly optimizes them using clinically verifiable rewards. Experiments on the MIMIC-CXR and IU X-ray datasets show that MARL-Rad consistently improves clinical efficacy metrics such as RadGraph, CheXbert, and GREEN scores, achieving state-of-the-art clinical efficacy performance. Further analyses show that MARL-Rad improves laterality consistency and produces more accurate and detailed reports. A blinded clinician evaluation further suggests that MARL-Rad produces reports clinically comparable to ground-truth reports.

Kaito Baba, Risa Kishikawa, Satoshi Kodera

Department of Cardiovascular Medicine The University of Tokyo Hospital, Tokyo, Japan

Radiology Report GenerationMultimodal AgentsRL for LLM AgentsLLM Agent Workflow OptimizationMulti-Agent Reinforcement Learning
AskPDF
May 30, 2026

A Machine-to-Machine Knowledge-Guided LLM Agent for Generalizable Radiotherapy Treatment Planning

In this work, we propose a prototype machine-to-machine (M2M) knowledge-guided Large Language Model (LLM) framework for automated radiotherapy treatment planning. In the proposed paradigm, Treatment Planning Parameter (TPP) distribution knowledge discovered by a Deep Reinforcement Learning (DRL) agent is transferred to an LLM agent through in-context learning, enabling autonomous iterative planning without human intervention. While standard LLM-based planning often lacks physical intuition and struggles with convergence, the integration of DRL-derived guidance constrains the agent to a physically valid parameter space. Experimental evaluations are performed across three diverse planning scenarios: basic prostate cases, complex prostate configurations with increased organ-at-risk (OAR) constraints, and liver cases. The evaluation results demonstrate that the guided LLM agent consistently achieves optimal planning scores while significantly reducing the number of iterations compared to unguided planning. Analysis of the final TPP configurations reveals that the agent successfully learns a hierarchical priority of objectives, effectively restoring a logical "cause-and-effect" relationship between parameter tuning and dosimetric outcomes. Crucially, this prototype framework exhibits robust generalizability, maintaining high planning quality regardless of specific patient anatomy, treatment site, or initial plan quality. By bridging the specialized optimization of DRL with the adaptive reasoning of LLMs, this M2M framework establishes a scalable foundation towards generalizable autonomous treatment planning, ultimately benefiting clinical practice in realistic environments.

Md Mainul Abrar, Xun Jia, Yujie Chi

Department of Physics, The University of Texas at Arlington, Arlington, 76019, TX · Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Baltimore, MD

LLM Planning
AskPDF