Click here to


Are you sure ?

Yes, do it No, cancel

Performance Evaluation of Deep Learning Methods Applied to CT Image Reconstruction

R Zeng1*, S Divel2 , Q Li1 , K Myers1 , (1) US Food and Drug Administration, Silver Spring, MD, (2) Stanford University, Stanford, CA,


(Sunday, 7/14/2019) 4:00 PM - 4:30 PM

Room: Exhibit Hall | Forum 9

Purpose: Deep learning (DL) methods applied to image reconstruction are attracting intensive interest in the field of CT. The purpose of this work is to evaluate the standard image quality metrics (CNR, MTF, NPS) and task-based performance metrics for characterizing images reconstructed with DL methods.

Methods: Two DL methods, one with three CNN layers (CNN3) and one with ten CNN layers (RED-CNN), were implemented and trained using the AAPM Low Dose CT Grand Challenge images. To train the networks, patches from simulated quarter-dose scans were processed and compared to their corresponding full-dose patient scans using the mean squared error as the loss function. To evaluate image quality, we simulated CT scans of the Catphan600 contrast module, a uniform water phantom, and the CCT189 LCD phantom to measure CNR, contrast-dependent MTF, NPS, and lesion detectability, respectively. FBP and a model-based iterative reconstruction method (MBIR) were also included in the evaluation for comparison.

Results: The preliminary results show that both DL methods suppressed noise and had better CNR than FBP. Similar to MBIR, these DL methods also had contrast-dependent MTF behavior. The NPS at the center of the images reconstructed with these DL methods was not rotationally symmetric, unlike FBP and MBIR. Finally, the task-based performance metric showed limited improvement of low-contrast detectability using DL methods compared to FBP.

Conclusion: Evaluation methods with standard phantoms add to our understanding of the behavior of DL methods in certain scenarios. Further work is needed to characterize CT reconstruction methods using DL, including studies making use of nonuniform phantom designs with challenging signals, realistic simulations, or clinical cases in order to more completely understand such methods as well as the relevance of standard phantom-based evaluation approaches.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email