Improving the Accuracy of Simultaneously Reconstructed Activity and Attenuation Maps Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Maximum-likelihood reconstruction of activity and attenuation (MLAA) with time-of-flight (TOF) is used for PET attenuation correction.
• MLAA faces challenges like crosstalk artifacts, slow convergence, and noisy attenuation maps (μ-maps).

Purpose of the Study:

• To develop and validate deep convolutional neural networks (CNNs) to overcome MLAA limitations for PET attenuation correction.
• To improve the accuracy and quality of attenuation maps in TOF-PET imaging.

Main Methods:

• Three CNN architectures (CAE, Unet, Hybrid) were trained to generate μ-maps from MLAA data using clinical brain PET/CT scans (40 patients).
• CNN-generated μ-maps were compared to CT-derived μ-maps (μ-CT) using Dice similarity coefficients.
• Activity concentration and binding ratios in specific brain regions were compared using reconstructed PET images.

Main Results:

• CNNs produced less noisy and more uniform μ-maps compared to MLAA, with better resolution of air cavities and bones.
• The Hybrid CNN network achieved high similarity to μ-CT (Dice: 0.79 bone, 0.72 air cavities), reducing quantification error to ~5%.
• Deep learning effectively mitigated crosstalk artifacts in MLAA reconstruction.

Conclusions:

• The proposed deep learning approach shows significant promise for accurate attenuation correction in TOF-PET systems.
• CNNs enhance the reliability and quantitative accuracy of PET imaging, particularly for challenging clinical cases like brain scans.