Bootstrap-Optimised Regularised Image Reconstruction for Emission Tomography.

Area of Science:

• Medical Imaging
• Computational Science
• Signal Processing

Background:

• Iterative image reconstruction in emission tomography often yields noisy images due to noisy data.
• Current regularization methods for noise compensation lack precise, automatic, and generalizable strength specification.
• Existing approaches include empirical approximations or computationally intensive cross-validation.

Purpose of the Study:

• To develop a novel, precise, and automatic methodology for optimizing regularization strength in iterative image reconstruction.
• To address the unresolved challenge of specifying regularization parameters for noisy emission tomography datasets.
• To improve the quality of reconstructed images in emission tomography without requiring ground truth information.

Main Methods:

• Proposes a new methodology embedded within iterative image reconstruction using bootstrapped replicates of measured data.
• Employs a conventional unregularized iterative update alongside a bootstrap replicate update.
• Optimizes regularization hyperparameters by fitting the regularized bootstrap update to the conventional measured data update.

Main Results:

• Demonstrates the method's effectiveness on positron emission tomography (PET) data across various noise levels.
• Achieves near-optimal image reconstructions in terms of reconstruction error.
• Eliminates the need for hyperparameter selection, tuning, or knowledge of ground truth or training data.

Conclusions:

• The novel bootstrapping approach provides precise regularization optimization for emission tomography image reconstruction.
• This method significantly enhances image quality and reduces noise without prior knowledge of the true image.
• Offers a computationally efficient and robust alternative to existing regularization techniques.