Deep Learning-Based Delayed PET Image Synthesis from Corresponding Early Scanned PET for Dosimetry Uptake Estimation

Area of Science:

• Medical Imaging
• Radiopharmaceuticals
• Artificial Intelligence

Background:

• Positron emission tomography (PET) imaging for radiopharmaceutical distribution is time-intensive due to multiple scans required for dosimetry.
• This necessitates significant time and inconvenience for participants.

Purpose of the Study:

• To develop a deep-learning model capable of generating delayed PET images from early-acquired images.
• To reduce the time and inconvenience associated with traditional multi-time-point PET imaging.

Main Methods:

• A generative adversarial network (GAN)-based paired image-to-image translation model was employed.
• Eighteen healthy participants received [18F]Fluorodeoxyglucose injections.
• The model generated delayed PET images from early scans, and standardized uptake value (SUV) means were compared to ground-truth data.

Main Results:

• The least square GAN and perceptual loss combination demonstrated optimal performance.
• Improved translation accuracy was observed as early image uptake times approached ground-truth times.
• SUV means for muscle, heart, liver, and spleen were estimated with reasonable accuracy.

Conclusions:

• Deep learning-based image-to-image translation is effective for generating functional PET images from other functional images in normal subjects.
• The model can predict organ-wise radiopharmaceutical activity in specific normal organs.
• This approach offers a potential solution to mitigate the time costs of dosimetry in PET imaging.