Author: Adnan Jafar, Xun Jia, Michael B. Roumeliotis 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Johns Hopkins University 🌍
Purpose: HDR brachytherapy (HDRBT) treatment planning is challenging due to the need for high-quality plans under time pressure, considering anatomy and applicator geometry. This study proposes an explainable virtual physician model that predicts the likelihood of approving a given treatment plan for cervical cancer patients undergoing tandem and ring-based brachytherapy and provides explainable and actionable suggestions for plan refinement.
Methods: A virtual physician model was developed based on an explainable convolutional neural network to predict the physician plans approval probability and a large language model (LLM) Llama 3.1 to generate planning guidance. The method takes distance histograms for the clinical target volume (CTV) and organs at risk (OARs) as well as the dose metrics used for plan assessment, and provides outputs as the likelihood of plan approval, along with suggestions for aspects of the plan that required improvement. The dataset included all cervical cancer cases at our institution between 2017 to 2024 that had at least 1 fraction of brachytherapy, treated with a tandem and ring (or ovoid) applicator, to train and test the proposed methodology. We evaluated the model’s performance by measuring accuracy, sensitivity, specificity, and area under the curve (AUC), as well as assessing its interpretability.
Results: The dataset included 33 cervical cancer cases. The proposed model can distinguish between clinically approved and unapproved plans, generated by perturbing the ground truth by 10-25% randomly, with an accuracy of 84%, a sensitivity of 86%, a specificity of 82%, and an AUC of 0.87. Additionally, the fine-tuned LLM can successfully generate plan improvement suggestions, achieving an accuracy of 92% when provided with prompts including EQD2 and Shapley values of current plans.
Conclusion: The virtual physician model can accurately predict the likelihood of plan approval with valuable interpretable guidance to treatment planning, which holds potential to substantially facilitate HDRBT treatment planning.