A Data-Efficient Unsupervised Deep Learning Deformable SPECT/CT Registration Framework for Voxel-Level Radionuclide Therapy Dosimetry
Abstract
Purpose
To develop a data-efficient, unsupervised deep learning framework for deformable SPECT/CT registration that supports voxel-based dosimetry in radionuclide therapy, particularly in clinical settings with limited imaging datasets.
Methods
We proposed NuFit-Reg, an unsupervised network for SPECT/CT registration using dual Swin-Transformer encoders with cross-stitch units for multimodal integration. To address data scarcity, a two-stage strategy was utilized: inter-patient pre-training on 58 iodine-131 scans, followed by few-shot fine-tuning on intra-patient sequential scans (12 patients, 3 time points) for temporal consistency. Performance was compared against Elastix and UTSRMorph using SSIM, wSSIM, LNCC, and MI, while deformation regularity was assessed via Jacobian determinants. Finally, the impact on voxel-level time–activity curve (TAC) fitting was analyzed to validate clinical utility.
Results
Compared with Elastix and UTSRMorph, it was demonstrated that NuFit-Reg achieved the highest overall registration accuracy, with a mean SSIM of 0.9523±0.0055. This advantage was most significant for the clinically-relevant weighted SSIM (0.8968 ± 0.0241), which exceeded Elastix by approximately 0.04, and remained consistent in repeated experiments. Importantly, NuFit-Reg reduced TAC errors, yielding an approximately 11.7% lower RMSE compared with Elastix for tumor voxels. In addition, NuFit-Reg completed SPECT/CT registration less than 9 s, while Elastix would take about 60 seconds using the same hardware device.
Conclusion
This study introduces NuFit-Reg, a two-stage unsupervised framework that achieves accurate and efficient deformable SPECT/CT registration under data-limited conditions. By combining inter-patient pre-training with intra-patient fine-tuning, NuFit-Reg provides a practical solution for supporting individualized voxel-based dosimetry in I-131 radionuclide therapy and potentially other radionuclide treatments in data-scarce clinical environments.