Optimizing Attenuation Correction in ⁶⁸Ga-PSMA PET Imaging Using Deep Learning and Artifact-Free Dataset Refinement.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Nuclear Medicine

Background:

• Quantitative accuracy in Positron Emission Tomography (PET) imaging relies on accurate attenuation correction (AC).
• Artifacts in CT-based AC (CT-AC) for 68Gallium-Prostate-Specific Membrane Antigen (68Ga-PSMA) PET, including motion and truncation, degrade image quality.
• These artifacts hinder the training of deep learning (DL)-based AC models.

Purpose of the Study:

• To develop and validate a novel artifact-refinement framework for PET-CT image datasets.
• To enable the training of an image-domain deep learning AC model using a purified dataset, eliminating the need for anatomical reference scans.
• To enhance the quantitative accuracy and robustness of 68Ga-PSMA PET imaging.

Main Methods:

• A residual neural network (ResNet) was trained on 828 whole-body 68Ga-PSMA PET-CT scans.
• Artifact-affected images were identified and excluded using voxel-level error metrics.
• The model was retrained on the refined dataset using an L2 loss function for improved performance.

Main Results:

• The model trained on the purified dataset showed significantly improved performance metrics (e.g., ME = -0.009 ± 0.43 SUV, MAE = 0.09 ± 0.41 SUV, SSIM = 0.96 ± 0.03).
• Enhanced quantitative accuracy and robustness were observed during internal and external validation compared to a model trained on unfiltered data.
• The artifact-refinement framework effectively mitigated errors caused by corrupted images.

Conclusions:

• The proposed data purification framework substantially improves deep learning-based AC for 68Ga-PSMA PET.
• This method enhances image fidelity and clinical applicability by enabling reliable PET imaging without anatomical references.
• The framework addresses critical challenges in artifact reduction for quantitative PET analysis.