DEMIST: A deep-learning-based task-specific denoising approach for myocardial perfusion SPECT

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Myocardial perfusion imaging (MPI) SPECT requires methods to process low-dose images for improved defect detection.
• Current denoising techniques may not optimize for observer performance in clinical tasks.

Approach:

• Developed DEMIST, a detection task-specific deep learning approach for denoising MPI SPECT images.
• DEMIST preserves crucial features for observer performance while reducing noise.
• Evaluated DEMIST using retrospective clinical data (N=338) at low radiation doses (6.25%-25%).

Key Points:

• DEMIST significantly improved the area under the receiver operating characteristics curve (AUC) compared to low-dose images and a task-agnostic DL method.
• Performance gains were consistent across patient sex and defect types.
• DEMIST enhanced image visual fidelity, assessed by RMSE and SSIM.

Conclusions:

• DEMIST demonstrates strong potential for denoising low-count MPI SPECT images.
• The approach improves observer performance in detecting perfusion defects.
• Further clinical evaluation of DEMIST is warranted for widespread adoption.