ESTRO 2025 - Abstract Book

S3425

Physics - Machine learning models and clinical applications

ESTRO 2025

3448

Digital Poster A unified deep-learning framework for enhanced patient-specific quality assurance Hui Khee Looe, Philipp Reinert, Julius Carta, Björn Poppe University Clinic for Medical Radiation Physics, Medical Campus Pius Hospital, Carl von Ossietzky University,

Oldenburg, Germany Purpose/Objective:

Intensity-modulated radiation therapy necessitates thorough patient-specific quality assurance (PSQA). Traditional measurements with detector arrays are labor-intensive, while independent calculation-based methods lack the assessment of machine performance during delivery. This study introduces a novel unified deep-learning (DL) framework that can combine the strengths of both approaches. Material/Methods: As shown in Figure 1, the unified framework consists of a forward prediction model that uses plan parameters to predict the measured dose distributions and a backward prediction model that reconstructs these plan parameters from actual measurements. To ensure broad generalization, the models were pre-trained using an extensive synthetic training dataset comprising 400,000 pairs of IMRT segments and the corresponding measurements generated based on mathematical models describing the physical processes of radiation transport and interaction within the medium and detector. These models were fine-tuned with an actual measured dataset of 400 selected synthetic IMRT segments to capture the individual machine-specific characteristics. Additional measurements of five clinical IMRT plans were used as the unseen test dataset. This method has been tested with an OD 1600 SRS and an OD 1500 detector array (PTW Dosimetry, Freiburg) with distinct spatial resolution and detector arrangement in combination with a dedicated upsampling model for the latter.

Figure 1: The proposed unified framework of PSQA that involves the implementation of a forward and a backward DL model.

Results: The forward model reached median gamma passing rates higher than 94% for all CP in the test plans measured using both detector arrays. The 3D gamma passing rates (2 mm / 2%) from comparing patients' original and recalculated dose distributions using the predicted actual plan parameters from the backward model lie between 94.7% and 98.5%. The DVH of the original and the recalculated dose distributions in both the TPS (Monaco, Elekta) and the Monte Carlo second-check software (VERIQA, PTW Dosimetry) showed no clinically relevant deviations, as presented exemplarily in Figure 2 for one clinical treatment plan (brain) in the test dataset.