Few shots transfer learning for universal SPECT denoising across diverse acquisition protocols

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accelerated Single Photon Emission Computed Tomography (SPECT) imaging enhances workflow efficiency but increases image noise.
• Deep learning methods for noise reduction require large datasets, hindering practical application for diverse acceleration protocols.

Purpose of the Study:

• To develop and evaluate a transfer learning framework for reconstructing accelerated SPECT images.
• To address data scarcity challenges in deep learning for SPECT image reconstruction.

Main Methods:

• A U-Net-based reconstruction framework was implemented with three strategies: single model, base model, and transfer model.
• SPECT bone scans from 103 patients were acquired under standard and five accelerated protocols.
• Quantitative metrics (PSNR, SSIM) and clinical evaluations assessed image quality and diagnostic performance.

Main Results:

• The transfer model demonstrated superior quantitative performance (highest PSNR 48.02, SSIM 0.9918) across all acceleration protocols.
• Clinical evaluations showed the transfer model achieved high scores for image quality, radionuclide detail, and diagnostic confidence.
• The transfer model's performance surpassed full-scan results in most evaluated metrics.

Conclusions:

• The transfer learning framework effectively mitigates data scarcity and enhances SPECT image reconstruction across various acceleration scenarios.
• This strategy allows for protocol-specific optimization by leveraging shared features and fine-tuning.
• The framework shows promise for integration into fast SPECT workflows, ensuring reliable imaging across diverse clinical settings.