Deep learning-guided joint attenuation and scatter correction in multitracer neuroimaging studies.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Attenuation correction (AC) is crucial for accurate PET imaging, especially on systems lacking CT or transmission scans.
• Dedicated brain PET scanners and hybrid PET/MRI systems present unique AC challenges.
• Deep learning approaches offer a potential solution for direct AC in image-space.

Purpose of the Study:

• To evaluate a deep learning-based direct attenuation correction (DLAC) method for various neuroimaging radiotracers.
• To assess the quantitative accuracy of DLAC compared to conventional methods.
• To identify limitations and potential vulnerabilities of the DLAC approach.

Main Methods:

• A deep convolutional neural network was trained to synthesize attenuation-corrected PET (PET-DLAC) images from non-AC PET (PET-nonAC) images.
• Four radiotracers (¹⁸F-FDG, ¹⁸F-DOPA, ¹⁸F-Flortaucipir, ¹⁸F-Flutemetamol) were used in 180 brain PET scans.
• Quantitative accuracy was assessed against CT-based AC (CTAC) as the reference standard.

Main Results:

• DLAC demonstrated superior performance compared to segmented AC (SegAC) across all radiotracers.
• The DLAC approach achieved less than 9% absolute SUV bias for the investigated neuroimaging radiotracers.
• Despite overall accuracy, DLAC showed vulnerability to outliers, leading to local pseudo uptake and false cold regions.

Conclusions:

• Direct image-space AC using deep learning provides a quantitatively acceptable method for brain PET imaging without CT.
• The DLAC approach is efficient for various radiotracers used in molecular neuroimaging.
• Careful consideration of potential outliers is necessary to mitigate local quantitative bias in DLAC.