Data-efficient unsupervised deep learning deformable SPECT/CT registration framework for voxel-level radionuclide therapy dosimetry: validation using clinical <sup>131</sup>I DTC therapy data.
Authors
Affiliations (4)
Affiliations (4)
- School of Nuclear Science and Technology, Institute of Nuclear Medical Physics, University of Science and Technology of China, Hefei, 230026, Anhui, China.
- Department of Radiation Oncology, The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, 230001, Anhui, China.
- Department of Nuclear Medicine, The First Affiliated Hospital of University of Science and Technology of China, Hefei, 230026, Anhui, China.
- School of Nuclear Science and Technology, Institute of Nuclear Medical Physics, University of Science and Technology of China, Hefei, 230026, Anhui, China. [email protected].
Abstract
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. We propose Nufit-Reg, an unsupervised deep learning network for deformable SPECT/CT registration. The network employs dual Swin-Transformer-based encoders to process CT and SPECT images separately, with cross-stitch units enabling structured feature sharing and multimodal integration across hierarchical stages. To overcome data scarcity, a two-stage training strategy was adopted. In the first stage, inter-patient pre-training was performed using randomly paired iodine-131 SPECT/CT scans (a total of 58 scans) to learn general anatomical correspondences. In the second stage, few-shot fine-tuning was performed using intra-patient sequential SPECT/CT pairs derived from quantitative scans (12 patients, each with three time points), enabling the model to adapt to patient-specific temporal consistency while being optimized across multiple patients. Registration performance was evaluated against Elastix and UTSRMorph using image similarity metrics, including structural similarity (SSIM), weighted SSIM (wSSIM), local normalized cross-correlation (LNCC), and mutual information (MI), as well as deformation regularity assessed via Jacobian determinant analysis. The downstream impact of registration on voxel-level time-activity-curve (TAC) fitting and tumor dose distributions (D<sub>2%</sub>, D<sub>98%</sub>) was also quantitatively analyzed. Nufit-Reg achieved a superior weighted SSIM (0.8968 ± 0.0241). It significantly outperforming Elastix and CT UTSRMorph. In dosimetry analysis, the enhanced registration accuracy led to an 11.7% reduction in tumor fitting RMSE compared to Elastix. These improvements aligned with more stable dose distributions. This was evidenced by reduced D<sub>2%</sub> extremes (110.81 to 95.1 Gy), and higher median D<sub>98%</sub> 3.30 Gy. 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 for other radionuclide treatments in data-scarce clinical environments.