Deep learning-based whole-body PSMA PET/CT attenuation correction utilizing Pix-2-Pix GAN

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiology

Background:

• Sequential PET/CT scans are crucial for oncology patient follow-up.
• Radiation dosage limits the frequency of sequential PET/CT studies.
• Low-dose CT scans are used for attenuation correction in PET/CT.

Purpose of the Study:

• To develop an AI tool for generating attenuation-corrected PET (AC-PET) images from non-attenuation-corrected PET (NAC-PET) images.
• To reduce the need for low-dose CT scans in oncology patients undergoing treatment follow-up.
• To maintain quantitative accuracy and image quality in AI-generated AC-PET images.

Main Methods:

• A 2D Pix-2-Pix generative adversarial network (GAN) deep learning model was developed.
• Paired AC-PET and NAC-PET images from 302 prostate cancer patients were used for training, validation, and testing.
• Two normalization strategies (SUV-based and SUV-Nyul-based) were employed.
• Performance was assessed using NMSE, MAE, SSIM, PSNR, and ICC for SUV metrics.

Main Results:

• The AI model achieved median performance metrics of 13.26% NMSE, 3.59% MAE, 0.891 SSIM, and 26.82 PSNR in the test cohort.
• High correlation (ICC 0.88-0.89) was observed between original and AI-generated SUVmax and SUVmean.
• Factors like lesion location, density, and uptake influenced the relative error in generated SUV metrics (p < 0.05).

Conclusions:

• The Pix-2-Pix GAN model effectively generates AC-PET images with high correlation in SUV metrics compared to original images.
• AI-generated PET images demonstrate potential for clinical use.
• This AI tool can reduce reliance on CT scans for attenuation correction while preserving quantitative markers and image quality.