Generating PET Attenuation Maps via Sim2Real Deep Learning-Based Tissue Composition Estimation Combined with MLACF.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiophysics

Background:

• Positron Emission Tomography (PET) imaging requires accurate attenuation correction (AC) for quantitative analysis.
• Computed Tomography (CT) is typically used for AC, but this adds radiation dose and complexity.
• Developing CT-less AC methods is a significant goal in PET research.

Purpose of the Study:

• To present the first Sim2Real deep learning (DL) based approach for generating human head attenuation maps using only simulated PET data.
• To evaluate the feasibility of DL-based tissue composition estimation for CT-less attenuation correction in PET.

Main Methods:

• A DL model was trained on simulated PET data to estimate a four-channel tissue composition map (soft tissue, bone, cavity, background) from a 2D non-AC PET image.
• Attenuation maps were generated from the DL-derived tissue composition maps.
• The DL-based attenuation maps were used for scatter+random estimation and as initial estimates for Maximum Likelihood Attenuation Correction Factor (MLACF) refinement.

Main Results:

• The DL model demonstrated the ability to estimate overall anatomical structures in clinical brain PET data.
• While some inaccuracies in anatomical detail were noted, particularly in neck-side slices, the DL-based AC achieved quantitative accuracy comparable to CT-based AC.
• The combined DL and MLACF approach showed promise as a CT-less AC solution.

Conclusions:

• The proposed DL-based method offers a viable approach for CT-less attenuation correction in PET.
• Combining DL-based tissue estimation with MLACF refinement presents a promising strategy for improving PET imaging without CT.
• Further development is warranted to address limitations in anatomical detail estimation for broader clinical application.