PET/CT for Brain Amyloid: A Feasibility Study for Scan Time Reduction by Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Alzheimer's disease diagnosis relies on 18F-florbetaben (18F-FBB) PET/CT scans.
• Full-time scans are lengthy, posing challenges for patient comfort and resource utilization.
• Developing faster imaging methods is crucial for improving diagnostic accessibility.

Purpose of the Study:

• To develop a convolutional neural network (CNN) model using a residual learning framework.
• To predict full-time 18F-FBB PET/CT images from shorter scan durations.
• To evaluate the model's performance in accurately reconstructing image quality and amyloid load.

Main Methods:

• A retrospective study included 22 cognitively normal subjects, 20 with mild cognitive impairment, and 42 with Alzheimer's disease.
• Short-time (1-5 min) PET/CT data were used to train a CNN model with single-slice or 3-slice input.
• Model performance was assessed quantitatively (RMSE, PSNR, SUV ratio) and qualitatively, comparing predicted to ground-truth full-time scans.

Main Results:

• Quantitative metrics improved with increased scan time; the 3-slice CNN model demonstrated better performance.
• The 3-slice CNN model using at least 2 minutes of scan data achieved comparable quantitative prediction to full-time scans.
• Qualitative analysis showed adequate to excellent image quality with the 3-slice CNN model, outperforming the single-slice model.

Conclusions:

• A 3-slice CNN model with a residual learning framework effectively predicts full-time 18F-FBB PET/CT images from short scans.
• This approach offers a promising method for reducing scan times in Alzheimer's disease diagnostics.
• The model provides accurate amyloid load quantification and high image quality, enhancing diagnostic efficiency.