Evaluation of data uncertainty for deep-learning-based CT noise reduction using ensemble patient data and a virtual...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Deep learning-based image reconstruction and noise reduction (DLIR) are increasingly used in clinical CT.
• Assessing data uncertainty is crucial for understanding DLIR stability against noise.

Purpose of the Study:

• Evaluate data uncertainty of a DLIR method using real patient data and a virtual imaging trial.
• Compare DLIR uncertainty with filtered-backprojection (FBP) and iterative reconstruction (IR).

Main Methods:

• Generated noise realizations using realistic projection domain noise insertion.
• Investigated impact of varying dose levels and denoising strengths on a ResNet-based deep convolutional neural network (DCNN).
• Utilized a virtual imaging trial framework with real patient data.

Main Results:

• DCNN uncertainty maps displayed more detailed structures than IR.
• DCNN bias maps showed less structural dependency, indicating higher sensitivity to input changes.
• Hotspots of DCNN uncertainty correlated with potential truth distortion but also improved small structure detection.

Conclusions:

• DCNN exhibits distinct uncertainty properties compared to IR, with implications for clinical stability and performance.
• The sensitivity of DCNN to input variations warrants careful consideration in clinical applications.
• DLIR methods like DCNN show promise for enhancing detection of small structures in CT imaging.