Image quality evaluation in deep-learning-based CT noise reduction using virtual imaging trial methods:...

Area of Science:

• Medical Imaging Physics
• Computational Imaging
• Radiology

Background:

• Deep-learning-based image reconstruction and noise reduction (DLIR) methods are increasingly used in clinical CT.
• Assessing DLIR performance is challenging as phantom studies may not reflect real-world patient data outcomes.
• DLIR models are primarily trained on patient images, necessitating patient-data-driven evaluation.

Purpose of the Study:

• To develop a patient-data-based virtual imaging trial framework for evaluating DLIR methods.
• To apply this framework to measure the spatial resolution properties of a specific DLIR technique.

Main Methods:

• Simulated lesions and noise were inserted into patient projection data at various contrasts and dose levels.
• Deep convolutional neural network (DCNN) and other reconstruction methods processed the data.
• Ensemble averaging of multiple noise realizations was used to calculate mean lesion signals.
• Modulation transfer functions (MTFs) were computed to assess in-plane and z-axis spatial resolution.

Main Results:

• DCNN spatial resolution degraded with lower contrast, reduced radiation dose, and increased denoising strength.
• At 25% dose and -10 HU contrast, DCNN showed a 59.5% in-plane and 4.1% z-axis MTF reduction compared to FBP.
• Spatial resolution ranking generally followed: FBP > DCNN-Weak > IR > DCNN-Medium > DCNN-Strong, especially at lower contrast/dose.

Conclusions:

• A novel patient-data-based virtual imaging trial framework was successfully developed and applied.
• The DCNN method, like other non-linear techniques, reduced spatial resolution under low contrast, low dose, and high denoising conditions.
• This framework provides a robust method for evaluating DLIR performance using patient-specific data.