Short-Axis PET Image Quality Improvement by Attention CycleGAN Using Total-Body PET

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Positron emission tomography (PET) imaging quality depends on scanner sensitivity and axial field of view (FOV).
• Conventional short-axis PET scanners with limited FOV (200-350 mm) yield reduced image quality during fast scans, impacting diagnostic reliability.
• Hardware limitations of short-axis PET scanners necessitate advanced solutions for improved image quality.

Purpose of the Study:

• To develop and evaluate a supervised deep learning model, CycleAGAN, for enhancing PET image quality from limited FOV scanners.
• To overcome hardware limitations and improve the diagnostic reliability of short-axis PET imaging.
• To maintain spatial consistency and improve image quality compared to existing methods.

Main Methods:

• A supervised deep learning model, CycleAGAN (based on CycleGAN), was proposed, incorporating an attention mechanism in the generator.
• The model focused on channel and spatial representative features, utilizing paired data for supervised learning to preserve spatial consistency.
• A dataset of 386 patients from Henan Provincial People's Hospital was prospectively collected, with training data from a total-body PET scanner (uEXPLORER).

Main Results:

• CycleAGAN demonstrated superior performance compared to traditional gray-level and learning-based methods, achieving the best results in Structural Similarity Index Measure (SSIM) and Normalized Root Mean Square Error (NRMSE).
• Expert ratings indicated that CycleAGAN produced images with distributions closest to the ground truth.
• The model effectively improved image quality for PET scanners with 320 mm FOV and showed good performance on scanners with shorter FOVs.

Conclusions:

• The proposed CycleAGAN model significantly enhances PET image quality from limited FOV scanners.
• This deep learning approach offers a viable solution to overcome hardware limitations, improving diagnostic accuracy and patient care.
• Integration of CycleAGAN into computer-aided diagnosis (CAD) systems can benefit both patients and radiologists.