Forward Model and Deep Learning Based Iterative Deconvolution for Robust Dynamic CT Perfusion.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Low-dose computed tomography (CT) perfusion maps exhibit poor signal-to-noise ratio, limiting their clinical utility.
• Existing methods like denoising and deconvolution or direct deep neural network (DNN) mapping have limitations in robustness to real-world data variations, particularly in stroke cases.

Purpose of the Study:

• To develop an iterative framework that directly generates high-quality perfusion maps from low-dose CT (LD-CT) dynamic data.
• To enhance the robustness of DNN-based methods by integrating the perfusion forward model and leveraging anatomical information.

Main Methods:

• An iterative framework combining the perfusion forward model with a DNN-based regularizer was proposed.
• Anatomical information from contrast-enhanced LD-CT images was utilized to train the DNN for mapping between low-dose and standard-dose perfusion maps, improving robustness.

Main Results:

• The proposed model successfully generates perfusion maps directly from LD-CT dynamic data.
• Empirical experiments demonstrated the model's qualitative and quantitative robustness to practical data perturbations.

Conclusions:

• The developed iterative framework offers a robust solution for generating accurate perfusion maps from LD-CT data.
• Leveraging anatomical information within the DNN improves the model's resilience to variations encountered in clinical applications like stroke imaging.