ESTRO 2024 - Abstract Book

S3785

Physics - Image acquisition and processing

ESTRO 2024

Results:

All image quality metrics showed significant improvement in the retrained models relative to the pre-upgrade model. The MAE decreased from 298.46 ± 27.33 HU to 75.75 ± 9.74 HU and 66.34 ± 6.85 HU for the TL-05 model and the retrained model using 60 patients (p<0.05), as shown in figure 1. There was no significant difference observed between the retrained models. Retraining durations varied, taking approximately 2.75 hours for TL-05 and up to 19.75 hours for TL-30. Retraining the new model using 60 patients required two days. Statistically, no significant differences were found between all models in terms of dosimetric parameters or NTCP. However, there was a notable variance, especially in small OARs, such as the cricopharyngeal muscle, as shown in figure 2a. However, the evaluation of the ΔNTCP (figure 2b) indicated no significant impact despite the substantial deviations in small OARs. All models met the 95% threshold for 2%/2mm and 3%/3mm gamma criteria.