Non-parametric Bayesian deep learning approach for whole-body low-dose PET reconstruction and uncertainty assessment

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Positron emission tomography (PET) is crucial for oncology, but low-dose (LD) PET imaging faces challenges in image quality and radiation exposure.
• Deep learning (DL) shows promise for LD PET reconstruction, yet uncertainty assessment is vital for clinical adoption.

Purpose of the Study:

• To develop and evaluate NPB-LDPET, a DL-based non-parametric Bayesian framework for LD PET reconstruction and uncertainty quantification.
• To assess the framework's performance against a Monte Carlo dropout benchmark using the Ultra-low-dose PET Challenge dataset.

Main Methods:

• Utilized an Adam optimizer with stochastic gradient Langevin dynamics (SGLD) for posterior distribution sampling.
• Evaluated global reconstruction accuracy (SSIM, PSNR, NRMSE), local lesion conspicuity (MAE, local contrast), and uncertainty map clinical relevance.
• Correlated uncertainty measures with the dose reduction factor (DRF).

Main Results:

• NPB-LDPET demonstrated significantly superior global reconstruction accuracy (p < 0.0001).
• Achieved a 21% reduction in MAE and an 8.3% improvement in local lesion contrast (p < 0.0001).
• Showcased a stronger correlation between predicted uncertainty and DRF (r^2 = 0.9174 vs. 0.6144).

Conclusions:

• The NPB-LDPET framework offers improved accuracy and uncertainty assessment for LD PET imaging.
• It has the potential to enhance clinical decision-making by providing more informative reconstructions.
• This approach supports reduced radiation exposure in PET scans without compromising diagnostic quality.