Using domain knowledge for robust and generalizable deep learning-based CT-free PET attenuation and scatter...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Deep learning (DL) shows promise for CT-free PET imaging, replacing traditional CT-based corrections.
• A key challenge is DL's limited ability to handle diverse tracers and PET scanners.
• Existing methods struggle with the heterogeneity inherent in PET data acquisition.

Purpose of the Study:

• To develop a robust and generalizable DL method for CT-free PET attenuation and scatter correction.
• To address the limitations of current DL approaches in handling tracer and scanner variability.
• To improve the clinical translation potential of DL in PET imaging.

Main Methods:

• A novel approach integrating domain knowledge into DL by decomposing the problem into low-frequency and high-frequency domains.
• Learning anatomy-dependent attenuation correction in the low-frequency domain for robustness.
• Preserving anatomy-independent high-frequency texture during processing.

Main Results:

• The proposed method demonstrated effectiveness and robustness across various external imaging tracers and different scanners, even when trained on a single tracer and scanner.
• Successful attenuation correction was achieved by separating low-frequency anatomical information from high-frequency texture.
• The approach proved generalizable beyond the initial training conditions.

Conclusions:

• The domain decomposition strategy offers a robust and transparent DL solution for CT-free PET imaging.
• This method overcomes the heterogeneity bottleneck, enhancing DL's applicability in diverse PET scenarios.
• The developed DL approach holds significant potential for clinical translation in nuclear medicine.