Deep learning-based Dosimetry in Radionuclide Therapy: Is It Worth the Effort?

We propose a novel unified framework to perform whole-body voxel-level dosimetry taking into account patient-specific tissue heterogeneity and activity distribution using Monte Carlo (MC) simulations and deep learning algorithms. We extended the core idea of the voxel-scale MIRD dosimetry formalism previously validated for positron emitters used in diagnostic imaging (
¹⁸
F) to radionuclides with complex decay schemes used in therapy (
¹⁷⁷
Lu). In this context, we trained a model to predict the deposited energy distribution obtained from MC simulations (specific S-values), while two-paired input channels consist of density map and dose distribution kernel in soft-tissue (single S-value) are fed into the network. Transfer learning was applied using our previous
¹⁸
F model fine-tuned on
¹⁷⁷
Lu dataset. Accordingly, whole-body dose maps were constructed through convolving specific S-values into time-integrated activity distribution obtained from SPECT images. The Deep Neural Network (DNN) predicted dose map was compared with the reference (Monte Carlo-based) and two MIRD-based methods, including single-voxel S-value (SSV) and multiple-voxel S-value (MSV) approaches. The results demonstrated that DNN, MSV and SSV show a comparable performance against the MC approach in soft tissue. However, in small size heterogeneous boundaries (lumbar region), DNN outperformed other approaches achieving lower bias (4%) compared to MSV (26%) and SSV (30%) techniques.