ESTRO 2024 - Abstract Book

S4502

Physics - Machine learning models and clinical applications

ESTRO 2024

1826

Digital Poster

Beyond clinical introduction of AI models: a fast and simple way to improve DL IMPT plan quality

Ilse van Bruggen, Minke Brinkman-Akker, Johannes Langendijk, Stefan Both, Erik Korevaar

UMCG, Radiotherapy, Groningen, Netherlands

Purpose/Objective:

Deep learning (DL) models have potential to enhance the quality and efficiency of radiotherapy treatment planning. However, their clinical adoption remains limited, and especially model maintenance and model sustainability post introduction is a pressing challenge. This study aims to improve our clinical DL model for intensity-modulated proton therapy (IMPT) plans for oropharyngeal carcinoma (OPC) patients.

Material/Methods:

In our current clinical practice, all IMPT plans for OPC patients are generated by a DL model in RayStation (RaySearch Laboratories AB, Stockholm, Sweden). The DL model was trained using contours and doses of 60 OPC patients. For new patients, a robust mimicking optimization algorithm for setup and range uncertainties 3mm/3% using voxel based mimicking and 21 perturbed scenarios was then used to generate a machine deliverable plan from the predicted dose distributions. Subsequently, a dosimetrist performs up to two hours of post-processing per plan, including tasks like adding objectives, adjusting objectives and continuing optimization. We evaluated standard plans of the initial model (sDL1) and post-processed plans of the initial model (pDL1) in 10 patients, identifying common updates made during post-processing to further improve the quality of the treatment plans. The model was improved by incorporating these updates via model configuration. Therefore, these modifications occurred rather at the model level than at the patient level. Model configuration involved adjustments like target and organs at risk (OARs) prioritization weights and optimization rounds. We iteratively refined the model until the standard plans of the updated model (sDL2) were comparable with pDL1 plans across all then patients. This took approximately 4 working days and was conducted by a single expert planner. Then, both the sDL1 (initial model) and the sDL2 (updated model) generated plans for 20 patients to evaluate the model improvement. All sDL1 and sDL2 plans were evaluated for clinical target volume D98% voxel-wise minimum dose >94% (using 28- perturbed scenario dose evaluations), maximum dose in the voxel wise maximum dose <113%, OAR doses and normal tissue complication probability (NTCP) (grade 2 dysphagia and xerostomia) based on the Dutch patient selection protocol. A Wilcoxon signed rank test was employed to evaluate the impact of these changes between sDL1 and sDL2 plans.

Results:

The most common post-processing actions involved increasing coverage of CTVelective, relaxing maximum dose levels and reduce dose to the esophagus (table 1). Model configuration successfully achieved the desired changes in