Personalized Internal Dose Prediction Using Consistency Model

Le résumé fourni par la source

Accurate and personalized voxel-level dosimetry remains a significant challenge in radionuclide therapy. Although Monte Carlo simulations provide accurate and reliable dose estimation, considerable computational time and cost limit their clinical application. In this study, the generative diffusion framework of a consistency model (CM) was applied to generate whole-body voxel-level internal dose maps from 3D PET/CT images through one-step sampling. PET and CT images were adopted as the input of the CM to provide the required information of radioactivity and physical density distributions for internal dose calculation. Whole-body dose maps calculated using the multiple voxel S-value method with 24 tissue-specific voxel dose kernels derived using MCNP simulations were used as the ground truth for model training and validation. Quantitative evaluation showed a mean absolute percentage error of$0.0170 \pm 0.0122$and a mean structural similarity of$0.9891 \pm 0.0066$, reflecting the consistency between the generated and reference dose maps. Even in tissue heterogeneous regions, such as lung boundaries, relative percent errors between the generated dose maps and ground truth remained lower than 10 %. The results indicated that the CM accurately modeled the complex interaction between radionuclide distribution and patient anatomy, providing a rapid and clinically feasible solution for personalized internal dosimetry in targeted radionuclide therapy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.