Design and validation of a deep learning GAN for predicting ¹ ⁷ ⁷Lu dose voxel kernels using GATE/GEANT4 Monte Carlo and IDAC-Dose 2.1 in phantom dosimetry.

Authors

Abstract

The rapid evolution of deep learning (DL), particularly the emergence of generative adversarial networks (GANs) has demonstrated strong potential for generating high-resolution, voxel-level dosimetric data. However, GAN-based approaches have not yet been broadly applied to the generation of tissue-specific dose voxel kernels (DVKs) in internal dosimetry. This gap underscores a critical unmet need for methods that can deliver Monte Carlo-level accuracy while achieving substantially improved computational efficiency. This study aimed to design a GAN architecture to train models that can generate ¹⁷⁷Lu DVKs for internal radiation dosimetry comparable to Monte Carlo standards. A GAN model was trained with paired CT-derived density kernel maps (DKs; voxel size 0.24³ cm³, matrix 15 × 15 × 15) and their corresponding Monte Carlo-simulated ¹⁷⁷Lu DVKs. The GAN-generated DVKs were applied to SPECT images of spherical phantoms with varying volumes and initial ¹⁷⁷Lu activity concentrations to compute mean absorbed doses. Results were validated against full GATE Monte Carlo simulations and IDAC-Dose 2.1 application. The GAN model predicted ¹⁷⁷Lu DVKs for water, soft tissue, muscle, and cortical bone with accuracies of 98.2%, 98.3%, 96.9%, and 94.2%, respectively (p > 0.05). Comparisons of GAN-derived ¹⁷⁷Lu DVKs with both GATE Monte Carlo and IDAC-Dose 2.1 produced very strong correlations (Pearson's r > 0.99; p = 0.0001). Bland-Altman analysis showed narrow limits of agreement at the 95% confidence interval (CI: -0.59 Gy to 0.44 Gy for GATE MC vs. GAN DVK, and CI: -0.01 Gy to -0.21 Gy for IDAC-Dose 2.1 vs. GAN DVK), with the vast majority of points falling within these ranges. These findings confirm the robustness and reliability of the GAN-predicted DVKs when benchmarked against established dosimetric methods. GAN-generated ¹⁷⁷Lu DVK convolution provides a potentially fast and computationally efficient approach for internal dosimetry, with strong agreement to established methods in phantom studies. While the clinical impact of tissue-specific DVKs remains dependent on broader improvements in quantitative imaging and dose-response characterization, this approach represents a promising proof-of-concept toward more physically consistent and scalable voxel-level dosimetry.

Tags

Topics