Agentic Radiotherapy Planning: Automating Outer-Loop Tuning via TextGrad

The Challenge of Precision Radiation Therapy

Radiotherapy treatment planning is one of the most computationally intensive aspects of cancer care. The process requires balancing tumor targeting with healthy tissue sparing across thousands of beam angles and dose constraints. Traditionally, medical physicists manually adjust parameters through an iterative process called outer-loop optimization, where each iteration can take hours and requires expert oversight.

This manual tuning presents several challenges:

• Time intensity: A single plan can require 10-20 iterations
• Human variability: Outcomes depend heavily on physicist expertise
• Scalability limits: High-demand periods strain specialist availability

Enter Agentic LLMs and TextGrad

The emerging approach uses Large Language Models as autonomous agents to handle outer-loop tuning. Rather than replacing the entire planning pipeline, these agents focus on hyperparameter optimization - adjusting objectives, constraints, and weighting schemes that guide the inner optimization loops.

TextGrad, developed as a framework for textual gradient-based optimization, enables this by:

1. Representing parameters as text: Treatment objectives and constraints become structured prompts
2. Computing gradients through language: The LLM identifies which adjustments would improve plan quality
3. Iterative refinement: Each cycle produces better parameter configurations

This differs from traditional numerical optimization by working directly with the semantic structure of treatment goals.

Technical Implementation

The agentic system operates through coordinated phases:

Phase 1: Plan Evaluation The current radiotherapy plan is analyzed against clinical goals. TextGrad encodes the plan quality metrics into a structured assessment that the LLM can process.

Phase 2: Parameter Adjustment The LLM, given context about the treatment site and observed shortcomings, proposes adjustments to the outer-loop parameters. These might include modifying dose constraints for specific anatomical structures or adjusting priority weights between tumor coverage and normal tissue sparing.

Phase 3: Validation Loop The proposed changes are implemented in the planning system, and the resulting plan is evaluated. This feedback loop continues until clinical acceptability criteria are met.

Why This Matters for Medical AI

The significance extends beyond radiotherapy. This approach demonstrates:

• Agentic reasoning in high-stakes domains: Moving beyond simple task automation to strategic decision-making
• Textual gradients for continuous optimization: A new paradigm where language models guide parameter spaces
• Human-AI collaboration patterns: The physicist remains in the loop for final approval while delegating repetitive tuning

Early results suggest 40-60% reduction in manual iteration time while maintaining plan quality metrics comparable to expert-planned treatments.

Broader Implications

As healthcare AI matures, we're seeing a shift from single-model solutions to coordinated agentic systems. TextGrad represents a bridge between symbolic reasoning (through text-based parameter manipulation) and neural optimization (through learned clinical intuition). This hybrid approach may prove particularly valuable in domains where safety constraints cannot be easily encoded numerically.

The success of outer-loop automation in radiotherapy could extend to other complex medical planning problems - radiation oncology, dosimetry, and even surgical planning where multiple competing objectives must be balanced.

Technical Resources

• TextGrad: Textual Gradient-Based Language Model Optimization - Microsoft research framework for gradient-based optimization using LLMs
• DoseOptimization.jl - Julia package for radiation therapy dose optimization
• Pinnacle^3 Planning System - Commercial treatment planning system used in many clinics
• OpenKBP - Open dataset for kidney tumor radiotherapy planning

Looking Forward

The next frontier involves expanding beyond single-parameter tuning to multi-agent coordination. Imagine one agent specializing in tumor coverage, another in cord sparing, with a meta-agent coordinating their competing recommendations. The text-based interface makes this coordination more interpretable than opaque numerical exchanges.

As these systems mature, the question becomes not just whether they can match human experts, but whether they can exceed human consistency while remaining fully auditable. In radiotherapy, where fractions of a Gray matter, having an AI system that can explain its reasoning in clinical language may be as important as the plan it produces.

The convergence of agentic reasoning and medical optimization represents a subtle but significant evolution in how we think about AI in healthcare - not as a replacement for expertise, but as a way to amplify and standardize it.