Learned Primal Dual Reconstruction for PET.

Area of Science:

• Medical Imaging
• Computational Science

Background:

• Positron Emission Tomography (PET) imaging requires robust algorithms for accurate image reconstruction.
• Existing methods like Maximum Likelihood Expectation Maximisation (MLEM) have limitations in noise reduction and detail recovery.

Purpose of the Study:

• To adapt, implement, and train the Learned Primal Dual (LPD) algorithm for PET image reconstruction.
• To evaluate the performance of LPD against MLEM and assess different training strategies.

Main Methods:

• Implementation and training of the LPD algorithm based on Adler and Öktem's work.
• Comparative analysis of LPD reconstructions with MLEM reconstructions using PET scanner data.
• Evaluation of LPD's robustness and performance across various noise levels and image features.

Main Results:

• LPD demonstrated superior performance in activity concentration recovery and noise reduction compared to MLEM when training data adequately represented noise levels.
• The LPD algorithm exhibited robustness against artifacts, even with novel features in the test data.
• Trained LPD models achieved image reconstruction in seconds.

Conclusions:

• The Learned Primal Dual algorithm offers a promising, efficient, and robust alternative for PET image reconstruction.
• Effective training strategies are crucial for optimizing LPD performance in PET imaging.