Enhancing bone scan image quality: an improved self-supervised denoising approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Nuclear Medicine

Background:

• Bone scans are crucial for skeletal lesion assessment but face limitations with low sensitivity and high noise in gamma camera imaging.
• Deep learning (DL) offers potential for image quality enhancement without increasing radiation or scan time.
• Existing self-supervised denoising methods like Noise2Noise (N2N) may introduce deviations from clinical standards in bone scans.

Purpose of the Study:

• To propose an improved self-supervised denoising technique for bone scans.
• To minimize discrepancies between DL-based denoised images and full scan images.
• To evaluate the clinical utility of DL denoising for reduced-time bone scans.

Main Methods:

• Retrospective analysis of 351 whole-body bone scan datasets.
• Utilized Noise2Noise (N2N), Noise2FullCount (N2F), and interpolated N2N (iN2N) denoising models.
• Trained networks on reduced scan times (5-50%) and mixed datasets; performed quantitative and clinical evaluations.

Main Results:

• DL denoising models generated images resembling full scans; iN2N closely mirrored full scan patterns.
• Quantitative analysis indicated improved denoising with longer input times and mixed count training.
• Clinical evaluation favored N2N and iN2N for resolution, noise, blurriness, and findings in quarter-time scans.

Conclusions:

• The improved self-supervised denoising technique effectively enhances bone scan image quality, minimizing deviations from clinical standards.
• The method shows promise for quarter-time scans without compromising diagnostic performance.
• This approach can improve bone scan interpretations and aid in accurate clinical diagnoses.