Synthesizing PET/MR (T1-weighted) images from non-attenuation-corrected PET images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Positron emission tomography (PET) imaging often requires complementary CT or MR scans for anatomical reference and attenuation correction (AC).
• Repeated PET/CT or PET/MR scans increase patient radiation exposure, cost, and time, particularly during treatment monitoring.
• Accurate quantitative PET imaging relies on reliable AC maps derived from CT or MR data.

Purpose of the Study:

• To develop a deep learning-based method for generating synthetic attenuation-corrected PET (sAC PET) and synthetic MR (sMR) images directly from non-attenuation-corrected PET (NAC PET) images.
• To reduce the necessity of additional anatomical imaging (CT/MR) during serial PET scans for patient monitoring.
• To mitigate radiation dose, economic burden, and time constraints associated with traditional PET/CT and PET/MR protocols.

Main Methods:

• Utilized a Wasserstein generative adversarial network (WGAN) framework.
• The model first processed NAC PET images to remove noise and artifacts, generating sAC PET images.
• Subsequently, sMR images were synthesized from the generated sAC PET images.

Main Results:

• The generative model was evaluated on paired PET/MR images from 80 clinical patients.
• Generated sAC PET and sMR images demonstrated high qualitative and quantitative similarity to real AC PET and real MR images.
• The method successfully produced synthetic images without requiring additional CT or MR acquisitions.

Conclusions:

• The proposed deep learning method effectively generates high-quality sAC PET and sMR images from NAC PET data.
• This approach has significant potential to decrease the frequency of supplementary anatomical imaging in PET studies.
• The method can enhance diagnostic efficiency, reduce patient costs, and minimize radiation risks associated with CT scans.