Pre-training with PBPK-based Digital Twin Enhances Deep Learning for Predictive Dosimetry
Nuklearmedizin - NuclearMedicine · 01.03.2025 · 64 (01) · 59
DOI: 10.1055/s-0045-1804320 ↗Kurz gefasst
Fachartikel in der Zeitschrift Nuklearmedizin - NuclearMedicine, erschienen am 01.03.2025. Verfasst von M Kassar und 15 weiteren Autor:innen.
Abstract (Originalsprache)
Ziel/Aim: Targeted radiopharmaceutical therapy (RPT) is has emerged as an effective treatment for mCRPC. However, prediction accuracy in dosimetry remains limited due to the scarcity of available data, as dosimetry data is challenging to obtain. This study introduces a PBPK-pretrained AI model for improved organ-specific dosimetry forecasting, enhancing time-series prediction accuracy by incorporating PBPK modeling in pre-training. Methodik/Methods: Data from 23 mCRPC patients treated with 177Lu-PSMA I&T RLT were analyzed, using pre-therapy PET/CT scans to derive organ-based SUVs for kidney, liver, spleen, and salivary gland dosimetry. A PBPK-pretrained, Transformer-based model was developed for dosimetry prediction. To support model training, 3,000 virtual patient scenarios were generated using PBPK modeling with XCAT phantoms, simulating PET characteristics and ligand distributions across therapy cycles. Synthetic PET images and dose-kernel methods were then used to calculate post-therapy dosimetry at multiple time points and cycles. Ergebnisse/Results: PBPK pre-training significantly improved the accuracy of dosimetry forecasting compared to forecasts generated without PBPK pre-training. The PBPK-pretrained model achieved significantly lower MAPE in dosimetry forecasting for the first cycle, with values of 4.13%±3.24 for the liver, 5.32%±5.33 for the spleen, 2.23%±2.63 for the kidneys, and 2.47%±2.39 for the salivary glands. In contrast, models without PBPK pre-training had much higher MAPE values across all organs. Furthermore, our approach with PBPK pre-training may significantly reduce estimation uncertainty in 4 cycles’ dosimetry forecasting, achieving a MAPE of 11.32%±8.26 for the liver, 7.48%±6.53 for the spleen, 7.63%±6.98 for the kidneys, and 7.37%±7.59 for the salivary glands, compared to the LSTM-based model, which showed higher MAPEs across all organs. Schlussfolgerungen/Conclusions: This study concludes that integrating pharmacological and physiological insights into the pre-training presents a promising approach to enhancing the accuracy dosimetry time-series forecasting for RLT treatment planning. Publication History Article published online: 12 March 2025 © 2025. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany
Zitieren
APA
Kassar, M., Chen, Y., Hu, J., Hong, J., Xue, S., Gafita, A., Rauscher, I., D'alessandria, C., Seifert, R., Ziegler, S., Wendler, T., Navab, N., Afshar-Oromieh, A., Rominger, A., Eiber, M., & Shi, K. (2025). Pre-training with PBPK-based Digital Twin Enhances Deep Learning for Predictive Dosimetry. Nuklearmedizin - NuclearMedicine, 64(01), 59. https://doi.org/10.1055/s-0045-1804320
Vancouver
Kassar M, Chen Y, Hu J, Hong J, Xue S, Gafita A, et al. Pre-training with PBPK-based Digital Twin Enhances Deep Learning for Predictive Dosimetry. Nuklearmedizin - NuclearMedicine. 2025;64(01):59. doi:10.1055/s-0045-1804320
BibTeX
@article{Kassar2025,
author = {Kassar, M and Chen, Y and Hu, J and Hong, J and Xue, S and Gafita, A and Rauscher, I and D'alessandria, C and Seifert, R and Ziegler, S and Wendler, T and Navab, N and Afshar-Oromieh, A and Rominger, A and Eiber, M and Shi, K},
title = {Pre-training with PBPK-based Digital Twin Enhances Deep Learning for Predictive Dosimetry},
journal = {Nuklearmedizin - NuclearMedicine},
year = {2025},
volume = {64},
number = {01},
pages = {59},
doi = {10.1055/s-0045-1804320}
}
EUVIMED