Deep learning-based attenuation map generation for myocardial perfusion SPECT.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Single-photon emission computed tomography (SPECT) accuracy is limited by attenuation artifacts in systems without CT transmission scanning.
• CT transmission scanning improves SPECT accuracy but increases radiation dose and can introduce misregistration artifacts due to patient motion.

Purpose of the Study:

• To develop a deep learning approach for estimating SPECT attenuation maps directly from emission data.
• To overcome limitations of CT-based attenuation correction in SPECT imaging.

Main Methods:

• Utilized both photopeak and scatter window SPECT images as input for a deep fully convolutional neural network.
• Trained and tested the network on cardiac SPECT/CT images from 65 patients, using CT-based attenuation maps as ground truth.
• Evaluated standard and adversarial training strategies.

Main Results:

• Generated synthetic attenuation maps that were qualitatively and quantitatively consistent with CT-based maps (NMAE of 3.60% ± 0.85%).
• SPECT images reconstructed with synthetic attenuation maps showed high consistency with those corrected using CT-based maps (NMAE of 0.26% ± 0.15%).
• Localized errors in the left ventricle myocardium and blood pool were minimal.

Conclusions:

• Developed a deep convolutional neural network capable of estimating reliable SPECT attenuation maps from emission data alone.
• The proposed method facilitates accurate attenuation correction for SPECT-only scanners, particularly for myocardial perfusion imaging.
• Offers a potential solution to reduce radiation exposure and motion artifacts in SPECT imaging.