Data Management and Network Architecture Effect on Performance Variability in Direct Attenuation Correction via Deep...

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate attenuation correction (AC) is crucial for interpreting SPECT myocardial perfusion imaging (MPI).
• Dedicated cardiac systems often lack transmission imaging capabilities, complicating AC.
• Previous deep learning (DL) methods showed feasibility but suffered from patient-to-patient performance variability.

Purpose of the Study:

• To investigate methods for overcoming performance variability in direct AC for SPECT MPI.
• To develop and evaluate advanced DL networks and data management strategies for improved AC accuracy.

Main Methods:

• Compared U-Net and Wasserstein cycle GAN (WCycleGAN) for DL-based AC (SPECT_DL).
• Implemented a data management strategy using clustering in a lower-dimensional space to select training data.
• Quantitatively analyzed global and regional AC performance.

Main Results:

• Advanced DL networks improved global AC performance with limited data.
• Regional AC accuracy did not show significant improvement.
• Clustered training demonstrated potential benefits for effective DL model training.

Conclusions:

• Advanced DL networks enhance the global performance of direct AC in SPECT MPI.
• The proposed data management strategy shows promise for effective training and reducing variability.
• Further research is needed to improve regional AC accuracy and facilitate clinical translation.