Attenuation correction using deep learning for brain perfusion SPECT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Non-uniform attenuation correction using computed tomography (CT) enhances single-photon emission computed tomography (SPECT) image quality and quantification.
• Widespread adoption is limited by the requirement for SPECT/CT scanners.

Purpose of the Study:

• To construct a convolutional neural network (CNN) auto-encoder (AE) capable of generating attenuation-corrected SPECT images directly from non-attenuation-corrected (No-AC) images.
• To evaluate the accuracy of AE-generated attenuation-corrected (AE-AC) images compared to CT-based attenuation correction (CT-AC).

Main Methods:

• An AE was trained using CNNs on brain perfusion SPECT datasets (270 training, 60 validation, 30 testing cases).
• The AE learned an end-to-end mapping between No-AC and CT-AC images.
• Accuracy was assessed using peak signal-to-noise ratio (PSNR) and structural similarity metric (SSIM), with voxel-by-voxel and region-by-region analysis.

Main Results:

• AE-AC images demonstrated superior PSNR (62.2) and SSIM (0.9995) compared to Chang-AC images (PSNR 57.9, SSIM 0.9985) when benchmarked against CT-AC.
• Visual and statistical analyses confirmed good agreement between AE-AC and CT-AC images.

Conclusions:

• The proposed AE-AC method effectively generates highly accurate attenuation-corrected brain perfusion SPECT images.
• This AI-driven approach provides a promising alternative for SPECT attenuation correction, potentially broadening its clinical application.