Assessment of deep learning-based PET attenuation correction frameworks in the sinogram domain

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Accurate attenuation correction (AC) is crucial for quantitative Positron Emission Tomography (PET) imaging.
• Deep learning (DL) offers potential for improving AC methods, particularly in the sinogram domain.

Purpose of the Study:

• To investigate and compare various deep learning frameworks for PET attenuation correction in the sinogram domain.
• To evaluate the performance of different DL models using both time-of-flight (TOF) and non-TOF PET data.

Main Methods:

• Implemented DL models for direct AC sinogram estimation and attenuation correction factor (ACF) estimation.
• Compared models using TOF and non-TOF PET data, with and without prior scatter correction.
• Utilized a segmentation-based AC map as a reference and PET/CT AC as the ground truth.

Main Results:

• DL models using TOF information significantly outperformed non-TOF models and segmentation-based methods (max SUV bias: 6.5% for TOF DL vs. 9.5% for non-TOF DL vs. 14.0% for segmentation).
• Direct estimation of AC sinograms showed no sensitivity to scatter correction, unlike ACF estimation methods.
• Direct AC sinogram prediction for TOF PET requires multiple input/output channels, potentially increasing computational cost.

Conclusions:

• Deep learning-based attenuation correction in the sinogram domain, especially with TOF PET data, demonstrates superior performance.
• Direct AC sinogram estimation is robust to scatter correction, simplifying the workflow.
• While effective, TOF-based DL sinogram methods may require significant computational resources.