Feasibility of direct brain ¹⁸F-fluorodeoxyglucose-positron emission tomography attenuation and high-resolution...

Area of Science:

• Medical Imaging
• Radiochemistry
• Artificial Intelligence

Background:

• Brain 18F-fluorodeoxyglucose-positron emission tomography (PET) requires accurate attenuation correction (AC) for precise quantitative analysis.
• Traditional AC methods often rely on CT scans, introducing additional radiation exposure.
• Deep learning offers a potential alternative for developing novel AC techniques.

Purpose of the Study:

• To develop and evaluate three deep learning-based attenuation correction (AC) methods for brain 18F-fluorodeoxyglucose-positron emission tomography (PET).
• The methods include an indirect approach, a direct approach, and a direct approach with high-resolution correction (direct+HRC).
• To assess the precision and accuracy of these novel AC methods compared to CT-based correction.

Main Methods:

• Three deep learning models (U-net architecture) were developed for AC: indirect (using synthetic CT), direct, and direct+HRC.
• Patient data included cranial MRI, CT, and PET scans (n=27) and MRI/CT (n=53).
• Image quality and precision were evaluated using Normalized Mean Squared Error (NMSE) and Structural Similarity (SSIM).

Main Results:

• Visual inspection showed no significant difference between CT-based AC and the deep learning methods.
• NMSE values were 0.281×10⁻³, 4.62×10⁻³, and 12.7×10⁻³ for indirect, direct, and direct+HRC methods, respectively.
• The direct+HRC method achieved a high SSIM of 0.975, indicating excellent image similarity.

Conclusions:

• The direct+HRC deep learning method provides accurate attenuation correction for brain PET without the need for CT scans.
• This method offers high-resolution correction capabilities, eliminating the need for specialized correction software.
• Deep learning-based AC presents a promising, radiation-free alternative for quantitative PET imaging.