Task-Specific Deep Learning Denoising for Myocardial Perfusion SPECT

The patent changes the optimization parameter from fidelity-based criteria (e.g., pixel-wise error) to task-specific performance metrics (e.g., observer performance in defect detection). This is achieved by incorporating task-specific loss functions that directly optimize for clinical detection accuracy rather than mere image reconstruction fidelity, thereby preserving task-relevant information while maintaining image quality.

The patent implements feedback mechanisms where the deep learning model is trained using task-specific performance metrics as feedback signals. The model receives feedback on its ability to preserve task-specific information (such as defect detectability) and adjusts its parameters accordingly, creating a closed-loop optimization process that aligns image reconstruction with clinical objectives.

The patent converts the harmful effect of noise in low-dose images into a beneficial outcome by using task-specific deep learning methods that are optimized to preserve detection-relevant information. Instead of treating noise reduction as merely improving visual quality, the approach leverages the noisy low-dose input to produce images that maintain or enhance task-specific performance (defect detection), thereby turning the limitation of low dose into an opportunity for task-optimized reconstruction.

The patent changes the optimization objective from general image fidelity to task-specific performance metrics. By training the deep learning model with loss functions that directly measure detection performance (such as observer performance in identifying perfusion defects), the system achieves superior task-specific results at reduced radiation doses compared to traditional fidelity-based methods.