Deep learning-based CT-free attenuation correction for cardiac SPECT: a new approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Cardiac SPECT imaging commonly uses CT-based attenuation correction (CTAC) to reduce artifacts.
• CTAC has limitations including inaccuracies, SPECT-CT mismatch, and additional radiation exposure.
• Deep learning (DL) offers a promising solution for CT-free attenuation correction (AC).

Purpose of the Study:

• To develop and evaluate a novel DL-based CT-free AC method for cardiac SPECT.
• To improve the accuracy and reduce artifacts in cardiac SPECT imaging without CT.
• To offer a safer and more efficient alternative to traditional CTAC.

Main Methods:

• Developed a feature alignment attenuation correction network (FA-ACNet) using a 3D U-Net framework.
• Trained the network on 167 cardiac SPECT/CT studies, utilizing multi-scale feature alignment with adversarial and distance metric learning.
• Evaluated performance using Mean Square Error (MSE), Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Bland-Altman analysis.

Main Results:

• FA-ACNet demonstrated high performance with excellent MSE, SSIM, and PSNR values in both cross-validation and an independent testing set.
• The DL-based method showed improved accuracy compared to standard U-Net by effectively aligning multi-scale features.
• Semi-quantitative analysis showed good agreement between DL-based AC and CTAC for summed stress/rest scores, with minimal changes in perfusion categories.

Conclusions:

• A novel DL-based CT-free AC approach for cardiac SPECT was successfully developed.
• The FA-ACNet method significantly enhances CT-free AC performance by leveraging multi-scale features from NAC SPECT and CT.
• This approach presents a promising alternative for future DL-based AC strategies in cardiac imaging.