Generation of PET Attenuation Map for Whole-Body Time-of-Flight ¹⁸F-FDG PET/MRI Using a Deep Neural Network Trained...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Nuclear Medicine Physics

Background:

• Accurate attenuation correction is crucial for quantitative analysis in positron emission tomography (PET)/magnetic resonance imaging (MRI) hybrid systems.
• Current methods like the Dixon-based 4-segment approach have limitations in whole-body PET/MRI attenuation correction accuracy.
• Deep learning offers potential for improving image reconstruction and correction techniques in medical imaging.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based method for enhanced whole-body PET/MRI attenuation correction.
• To compare the performance of the proposed convolutional neural network (CNN) approach against existing methods, including the 4-segment and MLAA algorithms.
• To assess the accuracy of attenuation maps generated by the CNN in terms of noise reduction and bone structure identification.

Main Methods:

• A modified U-net convolutional neural network (CNN) was trained using whole-body 18F-FDG PET/CT data from 100 cancer patients.
• The CNN learned to predict CT-derived attenuation maps (μ-CT) from MLAA-estimated activity (λ-MLAA) and attenuation maps (μ-MLAA).
• Performance was evaluated by comparing CNN-generated attenuation maps (μ-CNN) against μ-CT (ground truth) and other methods using Dice similarity coefficients and pixel-wise correlations.

Main Results:

• The proposed CNN method generated less noisy attenuation maps and demonstrated superior bone identification compared to the MLAA algorithm.
• The average Dice similarity coefficient for bone regions between μ-CNN and μ-CT was 0.77, significantly higher than μ-MLAA (0.36).
• The CNN-based attenuation correction showed the best pixel-by-pixel correlation with CT-based results and reduced activity map differences.

Conclusions:

• The developed deep neural network provides a more reliable attenuation map for 511-keV photons compared to the standard 4-segment method in whole-body PET/MRI.
• This deep learning approach enhances the accuracy of attenuation correction, potentially improving quantitative accuracy and lesion detection in PET/MRI studies.
• The findings suggest a promising role for CNNs in advancing attenuation correction techniques for hybrid PET/MRI systems.