A tissue-informed deep learning-based method for positron range correction in preclinical 68Ga PET imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Positron range (PR) blurs PET images, limiting spatial resolution and quantitative accuracy, especially with high-energy positron emitters like Gallium-68 (68Ga).
• Accurate positron range correction (PRC) is crucial for precise quantification in PET imaging.

Purpose of the Study:

• To develop and validate a deep learning-based method for positron range correction using 3D residual encoder-decoder convolutional neural networks (3D RED-CNNs).
• To incorporate tissue-dependent anatomical information via a u-map-dependent loss function for improved PRC.
• To evaluate the performance of different 3D RED-CNN architectures against standard methods.

Main Methods:

• Three 3D RED-CNN architectures (Single-channel, Two-channel, DualEncoder) were trained on simulated PET data.
• Models incorporated tissue-dependent anatomical information using a u-map-dependent loss function.
• Performance was evaluated using metrics like MAE, SSIM, CR, and CNR in simulations and preclinical 68Ga mouse studies, compared to Richardson-Lucy-based PRC (RL-PRC).

Main Results:

• CNN-based PRC methods demonstrated up to 19% SSIM improvement and 13% MAE reduction compared to RL-PRC.
• The Two-Channel model achieved superior contrast recovery (97% agreement for lung activity vs. 77% for RL-PRC) and contrast-to-noise ratio.
• CNN models maintained stable noise levels, unlike RL-PRC which increased noise; the Two-Channel model showed improved tumor delineation and reduced spillover artifacts in preclinical data.

Conclusions:

• Deep learning-based positron range correction using 3D RED-CNNs effectively enhances PET image quality and quantitative accuracy, particularly for 68Ga.
• The proposed u-map-dependent loss function and specific CNN architectures show significant potential for improving clinical PET imaging.
• Future research will focus on domain adaptation and hybrid training for broader model generalization.