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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Cardiology

Background:

• Low-dose myocardial perfusion SPECT imaging is crucial for reducing patient radiation exposure.
• Deep learning (DL) methods show promise for denoising these images but may not always improve clinical task performance.
• Existing DL denoising often optimizes for image fidelity, not necessarily for observer-relevant information.

Approach:

• A novel DL-based denoising method was developed, incorporating principles from model observers and human visual systems.
• The approach prioritizes preserving information critical for detection tasks, specifically identifying perfusion defects.
• The method was evaluated using a retrospective study on anonymized clinical myocardial perfusion SPECT data.

Key Points:

• The proposed DL denoising method significantly improved the performance of detecting myocardial perfusion defects.
• Compared to standard low-dose images, the enhanced images led to better diagnostic accuracy.
• Task-specific information preservation is key to enhancing DL's impact on clinical SPECT tasks.

Conclusions:

• DL-based denoising can effectively improve observer performance in low-dose myocardial perfusion SPECT.
• Preserving task-specific information is a critical factor for successful clinical translation of DL in medical imaging.
• This work highlights a pathway for optimizing DL denoising for improved diagnostic capabilities in nuclear cardiology.