DEMIST: A Deep-Learning-Based Detection-Task-Specific Denoising Approach for Myocardial Perfusion SPECT

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Myocardial perfusion imaging (MPI) single-photon emission computed tomography (SPECT) is crucial for diagnosing heart conditions.
• Acquiring MPI SPECT images at lower radiation doses or shorter times is desirable but can compromise image quality and diagnostic accuracy.
• Existing denoising methods may not optimally preserve features critical for detecting perfusion defects.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based denoising method (DEMIST) for low-dose MPI SPECT images.
• To ensure the denoising process preserves features essential for observer performance in detecting perfusion defects.
• To objectively assess DEMIST's efficacy compared to low-dose images and standard denoising techniques.

Main Methods:

• A detection task-specific deep learning approach (DEMIST) was developed, integrating model-observer theory and human visual system principles.
• DEMIST was retrospectively evaluated on 338 anonymized clinical MPI SPECT studies across two scanners.
• Image quality and observer performance (using channelized Hotelling observer and AUC) were assessed at 6.25%, 12.5%, and 25% of standard dose levels.

Main Results:

• DEMIST-denoised images demonstrated significantly higher area under the receiver operating characteristic curve (AUC) compared to low-dose images and images processed with a task-agnostic deep learning method.
• Improved performance was consistent across different patient sexes and types of perfusion defects.
• Quantitative metrics (RMSE, SSIM) confirmed DEMIST's enhancement of visual fidelity while preserving diagnostically relevant features.

Conclusions:

• DEMIST effectively denoises low-count MPI SPECT images, leading to improved detection of perfusion defects.
• The method enhances image quality and observer performance without sacrificing crucial diagnostic information.
• DEMIST shows significant promise for clinical application in reducing radiation dose while maintaining diagnostic accuracy in MPI SPECT.