Automated IMRT Plan Generation

• Automated IMRT plan generation is an advanced process that replaces manual treatment planning with algorithmic frameworks to ensure reproducible target coverage and effective normal tissue sparing.
• It employs a range of techniques including convex optimization, evolutionary algorithms, deep learning, and reinforcement learning to optimize dose distributions and beam configurations.
• Iterative multi-criteria loops, GPU acceleration, and adaptive control systems enable these systems to deliver clinically robust, time-efficient, and deliverable radiation therapy plans.

Automated intensity-modulated radiation therapy (IMRT) plan generation encompasses a diverse ecosystem of algorithmic frameworks, optimization pipelines, and emerging agents designed to systematically replace the historically manual, labor-intensive treatment planning process in radiation oncology. The central objective is to produce deliverable IMRT plans that meet or surpass clinical standards on target coverage and normal tissue sparing, with minimal manual intervention, increased reproducibility, and efficiency suitable for growing patient volumes. Below, the structural foundations, methodologies, and evaluative paradigms of automated IMRT planning are systematically detailed and referenced.

1. Algorithmic Architectures and Modalities

Automated IMRT plan generation is realized via a spectrum of algorithmic frameworks, spanning classical convex optimization, evolutionary multiobjective algorithms, deep neural networks, high-level agent-based control, and hybrid machine learning–optimization strategies.

• Weighted-Sum Convex Optimization: Central to most pipelines is a quadratic penalty functional on beamlet intensities, subject to dose–volume objectives and machine constraints. For instance, the Eclipse TPS engine minimizes a sum of quadratic penalties over targets (PTVs) and OARs, with additional hinge or saturation terms for constraint enforcement (Yang et al., 12 Oct 2025, Gao et al., 21 Jan 2025).
• Hierarchical Multiobjective Evolutionary Algorithms (MOEA): Evolutionary strategies generate a diverse set of Pareto-optimal plans, hierarchically optimizing penalty weights and shape parameters, followed by deterministic convex optimization at the sub-problem level (Holdsworth et al., 2012).
• Reinforcement Learning Agents: Deep RL agents, such as those employing actor-critic with experience replay (ACER), are used to tune planning hyperparameters discretely through Markovian trial-and-error, delivering rapid convergence and robustness to input variability (Abrar et al., 1 Feb 2025). RL approaches have also been employed for direct beam-angle optimization in a fully automated fashion, achieving improvements in conformity indices relative to clinical defaults (Bao et al., 2023).
• LLM–Driven Control: The deployment of LLMs in a zero-shot setting—where the agent receives only general task priors, not case-specific data—enables direct orchestration of clinical TPS APIs, leveraging chain-of-thought inference and arithmetic feedback for iterative, interpretable plan refinement (Yang et al., 12 Oct 2025).
• Deep Learning–Based Dose and Fluence Prediction: End-to-end neural models predict 3D dose distributions or even full fluence maps directly from volumetric images and contours (e.g., 3D U-Net or Swin-UNETR), bypassing explicit inverse optimization in some cases and integrating with TPS for final deliverability (Mgboh et al., 10 Nov 2025, Dahiya et al., 2021).
• Dose Mimicking and Knowledge-Based Optimization: Predicted dose distributions from machine learning models are converted to deliverable IMRT plans using convex dose-mimicking quadratic programming, with structures such as the QuadLin model explicitly balancing fidelity to neural predictions versus protocol prescription (Yousefi et al., 2022, Szalkowski et al., 23 Feb 2024).

2. Iterative and Multi-Criteria Planning Loops

Automated IMRT pipelines are, by necessity, iterative and multi-criteria, employing either explicit loop architectures or inherently generating a trade-off surface.

• LLM/TPS Interactive Loop: In the LLM-zero-shot framework, each iteration performs metric extraction (including DVH endpoints and constraint deviations), prompt construction embedding prior knowledge and history, LLM-based constraint update, and optimization inside the TPS, looping until stopping criteria are met (e.g., convergence, hard constraint satisfaction, iteration cap) (Yang et al., 12 Oct 2025).
• AIRTP: Automated Iterative RT Planning performs iterative refinement of objectives and constraints, driven by programmatic analysis of DVH metrics (extracted via TPS scripting) and systematic adjustment using rules or model outputs. Each cycle involves OAR/target DVH evaluation, constraint adjustment, re-optimization, and clinical scorecard scoring (Gao et al., 21 Jan 2025).
• MCO and Pareto Surface Exploration: Multi-criteria optimizers (MCO) produce entire Pareto surfaces, allowing automated or interactive navigation between plans optimized for target conformity and OAR sparing. Approaches such as low-segment MCO-IMRT (Khan et al., 2014) and NC-POPS for noncoplanar beam sets (Huang et al., 2021) generate deliverable plans distributed along the physically permissible trade-off front.
• Bayesian and Meta-Optimization: Outer-loop optimizers (e.g., Bayesian Optimization or Parallel Nelder–Mead) treat objective function weights and constraints as hyperparameters to be tuned, measuring plan quality with composite meta-scores constructed from lexicographically tiered clinical indices (Wang et al., 2022, Huang et al., 2021).

3. Mathematical and Computational Formalisms

Automated planning systems leverage and extend a range of mathematical constructs:

• Composite Objective Functions: Virtually all platforms employ a convex quadratic objective,

$$F(\mathbf{x}) = \sum_{i\in\{\mathrm{PTV}\}} w_i (D_i(\mathbf{x}) - D_i^\mathrm{pres})^2 + \sum_{j\in\{\mathrm{OAR}\}} w_j \max(0, D_j(\mathbf{x}) - c_j)^2,$$

with $w_i, w_j$ determined by clinical protocol or outer-loop agent (Yang et al., 12 Oct 2025, Gao et al., 21 Jan 2025).

• Constraint Update and Trend Analysis: Formulas for constraint adaptation in iterative schemes are typically proportional-control style,

$$c_j^{(t+1)} = c_j^{(t)} - \eta_j^{(t)} (\mathrm{Obs}_j^{(t)} - G_j),$$

where $\eta_j^{(t)}$ encodes adaptive step sizes (Yang et al., 12 Oct 2025).

• Robust and Deliverable Aperture Optimization: Robust direct aperture optimization (RDAO) employs large-scale MILPs for aperture specification under motion uncertainty, using candidate plan heuristics to rapidly generate feasible solutions (Ripsman et al., 2021).
• GPU-Accelerated Solvers: Real-time deployment is enabled by mapping all major kernels (dose calculation, gradient computation, fluence update) onto GPUs, demonstrating 20–40× speedups compared to CPU (0908.4421).
• Deep Model Losses: For fluence/dose prediction, composite losses combine voxelwise MSE or MAE with differentiable surrogates for DVH point errors to more closely match clinical quality indices (Mgboh et al., 10 Nov 2025, Dahiya et al., 2021).

4. Plan Evaluation Metrics and Clinical Validation

Automated IMRT plan quality is quantified using standardized and protocol-driven indices:

• Target Conformity: Conformity index (CI), often defined as $\frac{V_{\mathrm{Pres}}}{V_{\mathrm{PTV}}}$ or Paddick-style ratios, measures agreement between prescribed and delivered dose coverage (Yang et al., 12 Oct 2025, Huang et al., 2021).
• Homogeneity: Homogeneity index (HI), typically $(D_2 - D_{98})/D_\mathrm{pres}$ or related forms, quantifies dose variation across the PTV (Yang et al., 12 Oct 2025, Khan et al., 2014).
• Organ-at-Risk Sparing: OAR sparing is evaluated using dose-at-volume (e.g., $D_{50}$, $D_{0.1\mathrm{cc}}$), mean dose, and protocol-specific volume thresholds (e.g., $V_{20\,\mathrm{Gy}}$) (Gao et al., 21 Jan 2025, Yousefi et al., 2022, Dahiya et al., 2021).
• Composite Clinical Scores: Some pipelines define multi-tiered composite scores, integrating HI, CI, OAR means, and spill indices with explicit weighting to mimic clinical lexicography (Huang et al., 2021).
• Gamma Analysis and QA: Physical deliverability is confirmed using gamma index analysis on phantom delivery, with passing rates above clinical standards (e.g., $\geq98\%$ at 3\%/2 mm) (Szalkowski et al., 23 Feb 2024).
• Comparative Outcomes: Quantitative improvements over clinical plans include reductions in OAR mean doses (e.g., rectum mean dose $30.2\rightarrow21.9$ Gy in NC-POPS (Huang et al., 2021)), tighter hot-spot control ($D_{\max}$ reductions), and lower monitor unit (MU) counts (Khan et al., 2014).

5. Workflow Integration, Generalizability, and Clinical Applicability

Automated IMRT pipelines demonstrate broad compatibility and deployment potential across disease sites, clinical TPSs, and clinical settings:

• Vendor and API Integration: Most frameworks interoperate directly with commercial TPSs such as Varian Eclipse (via ESAPI/PyESAPI and C#/Python), Elekta Monaco (via GUI emulation or scripting), and RayStation. This enables automated access to structure sets, dose matrices, DVH metrics, and optimizer calls (Yang et al., 12 Oct 2025, Gao et al., 21 Jan 2025, Ayala et al., 2018).
• Generalizability and Zero-Shot Operation: Several systems, notably LLM-driven and dose-mimicking pipelines, operate without site-specific fine-tuning; transfer to new anatomic sites or modalities (e.g., IMRT $\rightarrow$ VMAT or TomoTherapy) only requires updates to clinical goal tables and protocol priors (Yang et al., 12 Oct 2025, Szalkowski et al., 23 Feb 2024).
• Independence from Planner Expertise: Agent-based and scripting-driven solutions minimize intra- and inter-planner variability, standardizing plan quality, and supporting deployment in both academic and resource-limited environments (Gao et al., 21 Jan 2025, Ayala et al., 2018).
• Modular and Interactive Extension: MCO and real-time plan navigation paradigms support both fully automated decision-making and rapid human-in-the-loop plan adaptation, facilitating institution-specific trade-off tailoring and physician engagement (Zhang et al., 2021, Khan et al., 2014).
• Throughput and Time Savings: End-to-end runtimes are reduced from hours (manual) to $0.3\text{–}1$ h (AIRTP), minutes (GPU-enabled re-optimization), or near real time (RL-driven or LLM-agent) without loss of clinical deliverability (Gao et al., 21 Jan 2025, 0908.4421, Yang et al., 12 Oct 2025).

6. Limitations, Open Challenges, and Future Directions

Current automated IMRT methodologies do embody limitations and ongoing areas for investigation:

• Personalization and Instance-Specificity: Standardized scorecards and generic protocol objectives may not capture institution- or patient-specific preferences; enhancing personalization, especially in agent-guided paradigms, remains an open challenge (Gao et al., 21 Jan 2025).
• Robustness and Uncertainty: Robust optimization to account for inter-fraction anatomical variation (e.g., respiratory motion) is achieved via robust constraints in RDAO, but widespread integration in commercial IMRT pipelines is not yet universal (Ripsman et al., 2021).
• Hybridization and Surrogate Acceleration: Integration of fast surrogate dose models (e.g., deep dose predictors for prompt evaluation) and multi-agent or multi-criteria steering (combining LLM reasoning with physics-based surrogates) is proposed for workflow acceleration and nuanced trade-off control (Yang et al., 12 Oct 2025).
• Cross-Modality and Data Adaptation: Studies demonstrate that models trained solely on IMRT can be adapted to VMAT, TomoTherapy, or different institutional protocols via mimic-based constraints and flexible optimization targets; further validation across diverse populations and imaging standards is needed (Szalkowski et al., 23 Feb 2024).
• User-in-the-Loop and Regulatory Integration: For clinical translation, pipelines must be seamlessly embedded into QA, documentation, and physician review systems, with transparent decision logs and reproducible outputs (Yang et al., 12 Oct 2025, Gao et al., 21 Jan 2025).

7. Representative Workflows and Comparative Outcomes

The following table summarizes characteristic elements and findings for selected state-of-the-art automated IMRT pipelines:

Pipeline/Agent	Technical Core	Clinical Outcome Highlights
LLM/TPS Agent (Yang et al., 12 Oct 2025)	Zero-shot GPT-4.1 agent + ESAPI scripting	Dmax 106.5% vs 108.8% (clinical); improved CI
AIRTP (Gao et al., 21 Jan 2025)	Iterative scripting, RapidPlan, DL contours	Scorecard gains 10–20 pts over manual; runtime 0.3–1h
GPU QP (0908.4421)	Penalty-based QP on CUDA	20–40× CPU speedup; runtime ~2–3 s per plan
RL/ACER (Abrar et al., 1 Feb 2025)	Actor–Critic, DVH reward	>93% perfect ProKnow scores; robust to adversarial
Pareto MOEA (Holdsworth et al., 2012)	Hierarchical evolutionary MOO	Diverse Pareto front; clinical constraint satisfaction
Dose-Mimic + MCO (Szalkowski et al., 23 Feb 2024)	3-stage 3D U-Net + dose mimic + MCO	<5% deviation in all clinical goals vs predictions
QuadLin (Yousefi et al., 2022)	Convex dose-mimicking + prescription	+21% in criteria satisfaction over predicted dose

In sum, automated IMRT plan generation leverages multi-layered computational, statistical, and agent-based paradigms to achieve reproducible, high-quality, and efficient treatment planning, with current systems already matching or exceeding clinical practice across a broad range of protocols, case types, and institutions. Ongoing developments emphasize expanding personalization, uncertainty integration, cross-modality flexibility, and real-time clinical integration.