Human-like AI-based Auto-Field-in-Field Whole-Brain Radiotherapy Treatment Planning With Conversation Large Language Model Feedback

TL;DR

This work addresses WBRT planning bottlenecks due to manual hyperparameter tuning and iterative refinement by introducing a two-module AI pipeline: a Hyperparameter Prediction module that uses geometric features to automatically set Auto-FiF parameters in the RayStation TPS, and a Conversation module leveraging Whisper and GPT-4o to translate physician feedback into structured TPS adjustments. On 15 independent WBRT cases, DL-generated plans matched clinical plans with no significant dosimetric differences in $D1\%$ and $D99\%$ for the CTV and mean OAR doses, while achieving a total workflow around $7$ minutes and inference under $1$ minute for hyperparameters. Explainability via Integrated Gradients identified geometry-driven drivers of the predictions, and the Conversation module demonstrated practical plan refinements through natural-language feedback that improved conformity and hotspot control. The proposed end-to-end, one-click planning framework holds promise for accelerating WBRT workflows, enabling online adaptive radiotherapy, and generalizing to more complex sites, with potential gains in efficiency and consistency. $D95\%$ normalization was used to ensure fair dosimetric comparisons, and $p>0.05$ in reported tests supports clinical equivalence.

Abstract

Whole-brain radiotherapy (WBRT) is a common treatment due to its simplicity and effectiveness. While automated Field-in-Field (Auto-FiF) functions assist WBRT planning in modern treatment planning systems, it still requires manual approaches for optimal plan generation including patient-specific hyperparameters definition and plan refinement based on quality feedback. This study introduces an automated WBRT planning pipeline that integrates a deep learning (DL) Hyperparameter Prediction model for patient-specific parameter generation and a large-language model (LLM)-based conversational interface for interactive plan refinement. The Hyperparameter Prediction module was trained on 55 WBRT cases using geometric features of clinical target volume (CTV) and organs at risk (OARs) to determine optimal Auto-FiF settings in RayStation treatment planning system. Plans were generated under predicted hyperparameters. For cases in which the generated plan was suboptimal, quality feedback via voice input was captured by a Conversation module, transcribed using Whisper, and interpreted by GPT-4o to adjust planning settings. Plan quality was evaluated in 15 independent cases using clinical metrics and expert review, and model explainability was supported through analysis of feature importance. Fourteen of 15 DL-generated plans were clinically acceptable. Normalized to identical CTV D95% as the clinical plans, the DL-generated and clinical plans showed no statistically significant differences in doses to the eyes, lenses, or CTV dose metrics D1% and D99%. The DL-based planning required under 1 minute of computation and achieved total workflow execution in approximately 7 minutes with a single mouse click, compared to 15 minutes for manual planning. In cases requiring adjustment, the Conversational module successfully improved dose conformity and hotspot reduction.

Figures (7)

• Figure 1: Workflow of the Auto-FiF treatment planning for WBRT, augmented with an explainable DL-based Hyperparameter Prediction module and a Conversation module powered by an LLM.
• Figure 2: Architectures of the DL models for Auto-FiF hyperparameter prediction, including a model for target coverage priority, a model for the number of subfields, and the third one that reused the subfield classifier backbone with additional layers to predict minimum segment MU per fraction and minimum segment area. An explainable AI module was employed for attributing model predictions to geometric features.
• Figure 3: Dendrogram illustrating the hierarchical clustering of geometric features to identify feature relationships, guiding the iterative feature selection process for optimizing model performance.
• Figure 4: Conversation module for physician-in-the-loop planning. Verbal feedback from the physician was converted to text using Whisper, then interpreted by GPT-4o to produce plan improvement suggestions that align with the physician’s intent.
• Figure 5: Examples of clinically approved plans (top) vs plans generated by the Hyperparameter Prediction module (bottom) for approved (left), approved with minor edits (middle), and rejected (right) cases. The middle red region indicates the CTV.