Data-efficient unsupervised deep learning deformable SPECT/CT registration framework for voxel-level radionuclide...

Area of Science:

• Medical Imaging
• Radiotherapy
• Deep Learning

Background:

• Accurate image registration is crucial for quantitative SPECT/CT analysis in radionuclide therapy.
• Data scarcity in clinical settings poses a challenge for developing robust registration models.

Purpose of the Study:

• To develop a data-efficient, unsupervised deep learning framework for deformable SPECT/CT registration.
• To support voxel-based dosimetry in radionuclide therapy, especially with limited imaging datasets.

Main Methods:

• Proposed Nufit-Reg, an unsupervised deep learning network using dual Swin-Transformer encoders.
• Employed a two-stage training strategy: inter-patient pre-training and few-shot intra-patient fine-tuning.
• Evaluated registration accuracy using image similarity metrics and Jacobian determinant analysis, and assessed downstream dosimetry impact.

Main Results:

• Nufit-Reg achieved superior weighted SSIM (0.8968 ± 0.0241) compared to existing methods.
• Registration accuracy improvements led to an 11.7% reduction in tumor fitting RMSE.
• Demonstrated more stable dose distributions with reduced D2% extremes and higher median D98%.

Conclusions:

• Nufit-Reg offers accurate and efficient deformable SPECT/CT registration in data-limited conditions.
• The two-stage training approach effectively addresses data scarcity.
• Provides a practical solution for individualized voxel-based dosimetry in I-131 radionuclide therapy.