Synthetic Attenuation Correction Maps for SPECT Imaging Using Deep Learning: A Study on Myocardial Perfusion Imaging.

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• CT-based attenuation correction is crucial for accurate SPECT cardiovascular imaging.
• Many SPECT cameras lack CT scanners, especially in developing nations.
• Additional CT scans increase patient radiation dose and logistical challenges.

Purpose of the Study:

• To develop a deep learning model for generating linear attenuation coefficient maps from SPECT images.
• To assess the feasibility of using these synthetic maps for attenuation correction in myocardial perfusion SPECT.

Main Methods:

• A 2D U-Net deep learning model was trained on 312 myocardial perfusion SPECT studies (99mTc-sestamibi).
• Synthetic attenuation correction maps (ACMs) were generated for 66 test patients.
• Quality of synthetic ACMs and reconstructed SPECT images were evaluated using metrics and compared to standard data.

Main Results:

• High-quality synthetic ACMs achieved MSSIM of 0.97 and NMAE of 3.08%.
• Reconstructed SPECT images showed excellent quality with MSSIM of 0.99 and NMAE of 0.23%.
• Voxel-level agreement for reconstructed images was within [-9.04; 9.00]%, with minimal impact on clinical scores.

Conclusions:

• Deep learning effectively generates accurate attenuation correction maps from cardiac SPECT data.
• These synthetic ACMs are suitable for myocardial perfusion SPECT, potentially replacing additional CT scans.
• This approach enhances SPECT imaging accessibility and reduces patient radiation exposure.