Deep-learning-based estimation of attenuation map improves attenuation correction performance over direct attenuation...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Deep learning (DL) offers potential improvements for attenuation correction (AC) in myocardial perfusion (MP) SPECT.
• Optimizing and comparing DL-based direct and indirect AC methods is crucial, especially considering SPECT/CT image mismatch.

Purpose of the Study:

• To optimize and compare deep learning-based direct and indirect attenuation correction methods for myocardial perfusion SPECT.
• To evaluate the performance of these methods under conditions with and without SPECT/CT image mismatch.

Main Methods:

• Simulations using XCAT phantoms (n=100) and retrospective analysis of patient SPECT/CT scans (n=34) with 99mTc-sestamibi.
• Reconstruction using OS-EM with and without AC, modeling SPECT/CT mismatch.
• Optimization of a 3D conditional generative adversarial network (cGAN) for indirect (DL-estimated attenuation map) and direct (DL-generated AC SPECT) AC.

Main Results:

• SPECT/CT image mismatch negatively impacted DL-based AC performance.
• The indirect DL-based AC approach demonstrated superiority over the direct approach across various physical and clinical metrics, even when mismatch was present.

Conclusions:

• Deep learning-based estimation of attenuation maps for AC is a more robust and effective strategy compared to the direct generation of AC SPECT.
• The indirect DL-AC method shows significant advantages for improving MP SPECT quality and reliability.