Dose Reduction in Scintigraphic Imaging Through Enhanced Convolutional Autoencoder-Based Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Low-dose bone scintigraphy is crucial for patient safety but can compromise image quality.
• Enhancing low-dose images is essential to maintain diagnostic accuracy while reducing radiation exposure.
• Deep learning offers potential for image reconstruction and quality improvement in medical imaging.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach, the Enhanced Convolutional Autoencoder (ECAE), for enhancing low-dose bone scintigraphy images.
• To reduce patient radiation exposure without sacrificing diagnostic quality.
• To validate the ECAE model using quantitative metrics and expert qualitative assessments.

Main Methods:

• A supervised learning framework using paired low- and full-dose bone scintigraphy images from 105 patients.
• The ECAE architecture incorporates multiscale feature extraction, channel attention, and residual blocks.
• Model training and validation involved Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and expert evaluations.

Main Results:

• The ECAE model significantly improved PSNR and SSIM, especially at 30-70% of full dose.
• Expert evaluations confirmed enhanced anatomical visibility, noise reduction, and diagnostic detail preservation.
• Denoised images were preferred over original full-dose scans in 66% of blinded evaluations.

Conclusions:

• The ECAE model effectively reconstructs high-quality bone scintigraphy from reduced-dose acquisitions.
• This deep learning approach enables significant dose reduction in nuclear medicine.
• The method enhances patient safety, workflow efficiency, and has positive environmental implications.