Segmentation-Free PVC for Cardiac SPECT Using a Densely-Connected Multi-Dimensional Dynamic Network.

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Limited resolution in nuclear imaging causes partial volume effects (PVEs), degrading image sharpness and quantitative accuracy.
• Anatomical-guided partial volume correction (PVC) methods using CT/MRI are effective but require complex registration and segmentation.
• Obtaining accurate cardiac segmentation is challenging in SPECT imaging due to scanner limitations and motion artifacts.

Purpose of the Study:

• To develop a deep learning-based method for rapid cardiac SPECT PVC that eliminates the need for anatomical information and segmentation.
• To introduce a densely-connected multi-dimensional dynamic mechanism for adaptive kernel adaptation within the neural network.
• To incorporate intramyocardial blood volume (IMBV) as a clinical-relevant loss function for network optimization.

Main Methods:

• A novel deep learning network with a densely-connected multi-dimensional dynamic mechanism was developed for cardiac SPECT PVC.
• The network was trained and validated on 28 canine cardiac SPECT studies acquired with a hybrid SPECT/CT scanner.
• Intramyocardial blood volume (IMBV) was used as an additional loss function to optimize network performance.

Main Results:

• The proposed deep learning network achieved fast cardiac SPECT PVC without requiring anatomical information or segmentation.
• The densely-connected dynamic mechanism significantly improved the network's performance compared to a network without this feature.
• The method produced statistically comparable IMBV measurements to traditional anatomical-guided PVC methods.

Conclusions:

• Deep learning offers a viable solution for efficient cardiac SPECT PVC, overcoming limitations of anatomical-guided approaches.
• The developed network, featuring a dynamic mechanism and IMBV loss, demonstrates superior performance and clinical relevance.
• This approach holds significant potential for clinical translation, improving diagnostic accuracy in cardiac SPECT imaging.