Linearized Analysis of Noise and Resolution for DL-Based Image Generation.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Radiology

Background:

• Deep learning (DL) CT image generation methods typically use RMSE and SSIM for evaluation.
• Conventional model-based image reconstruction (MBIR) methods assess image properties like resolution and noise, often requiring time-consuming Monte Carlo (MC) simulations.
• Linearized analysis has been used for MBIR to characterize noise and resolution without MC simulations.

Purpose of the Study:

• To investigate the applicability of network linearization to DL networks for efficient characterization of resolution and noise.
• To enable physics-related image quality measures for DL applications without MC simulations.

Main Methods:

• Applied network linearization, inspired by MBIR techniques, to a DL network (FBPConvNet).
• Conducted extensive numerical evaluations using both computer simulations and real CT data.
• Developed a generic method for computing covariance images for network linearization, applicable to DL modules combined with linear operators like filtered-backprojection (FBP).

Main Results:

• Network linearization proved effective for characterizing image noise and resolution in DL CT reconstruction under normal exposure settings.
• The method successfully avoided the need for MC simulations for image property assessment.
• Provided computational tools for implementing network linearization.

Conclusions:

• Network linearization offers an efficient and accessible approach for evaluating image quality in DL-based CT reconstruction.
• The methodology is general and supports flexible compositions of DL modules and linear operators.
• This work facilitates the adoption of physics-based image quality metrics in DL imaging applications.