Comparison of deep learning-based emission-only attenuation correction methods for positron emission tomography

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Accurate attenuation correction is crucial for quantitative positron emission tomography (PET) imaging.
• Traditional methods often rely on separate computed tomography (CT) scans for attenuation correction.
• Developing methods using only emission PET data is desirable to reduce scan time and radiation dose.

Purpose of the Study:

• To compare two convolution neural network (CNN) approaches for attenuation correction using only emission PET data.
• To evaluate the performance of CNNs in generating attenuation coefficient (μ) maps.
• To assess the suitability of CNN-based methods for scatter estimation in MLAA reconstruction.

Main Methods:

• Two CNN approaches were developed: μ-CNN_NAC (generating μ-maps from non-attenuation-corrected PET images) and μ-CNN_MLAA (improving μ-maps from MLAA reconstruction).
• A combined approach, μ-CNN_MLAA+NAC, was also investigated.
• Training and testing utilized PET/CT scans from 100 ¹⁸F-FDG and 50 ⁶⁸Ga-DOTATOC datasets.

Main Results:

• μ-CNN_NAC showed a 2.5% error in scatter estimation but over 7% error in attenuation correction factors and overestimated lung μ-values.
• The μ-CNN_MLAA+NAC approach yielded the best results in recovering fine bone structures.
• Activity images corrected with μ-CNN_MLAA and μ-CNN_MLAA+NAC showed superior similarity to CT-based reconstructions, with minimal errors for lung and bone cancer lesions.

Conclusions:

• μ-CNN_NAC is feasible for scatter estimation, addressing the 'chicken-egg' dilemma in MLAA reconstruction.
• μ-CNN_MLAA outperformed μ-CNN_NAC in overall accuracy.
• The combined μ-CNN_MLAA+NAC approach offers the most comprehensive solution for accurate attenuation correction and lesion quantification in PET imaging.