Towards prediction of pseudo-normal SPECT image data using variational autoencoder.

Area of Science:

• Medical Imaging
• Radiochemistry
• Artificial Intelligence

Background:

• Single Photon Emission Computed Tomography (SPECT) imaging can produce abnormal results.
• Harmonizing SPECT data is crucial for accurate analysis.
• Generating pseudo-normal SPECT data is a novel approach.

Purpose of the Study:

• To assess the feasibility of creating pseudo-normal SPECT data from abnormal images.
• To develop an on-the-fly data harmonization technique using pseudo-normal images.
• To evaluate the performance of a Variational Autoencoder (VAE) for this task.

Main Methods:

• A VAE model was developed to process 2D sinograms of brain SPECT ([123I]-FP-CIT).
• The VAE was trained on simulated SPECT data derived from MR scans with varying uptake levels.
• Performance was measured using Dice Similarity Coefficient (DSC) and specific binding ratio.

Main Results:

• The VAE achieved a mean DSC of 80% for the left basal ganglia and 84% for the right.
• The model demonstrated high consistency in predicting basal ganglia shape (DSC coefficient of variation < 1.1%).

Conclusions:

• VAE effectively estimates individualized pseudo-normal radiotracer distribution from abnormal SPECT images.
• This method shows promise for harmonizing SPECT data.
• Limitations include limited real MR data and a simplified simulation setup.