Contrast- and noise-dependent spatial resolution measurement for deep convolutional neural network-based noise...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Processing

Background:

• Deep convolutional neural networks (DCNNs) are used for noise reduction in clinical computed tomography (CT).
• Accurate assessment of DCNN spatial resolution is crucial for clinical application.
• Traditional phantom-based measurements may not reflect DCNN performance on patient data.

Purpose of the Study:

• To propose and validate a patient-data-based framework for measuring the spatial resolution of DCNN noise reduction methods in CT.
• To investigate the impact of lesion contrast, radiation dose, and DCNN denoising strength on spatial resolution.

Main Methods:

• Developed a framework involving lesion and noise insertion in the projection domain.
• Utilized lesion ensemble averaging and modulation transfer function (MTF) measurement from an oversampled edge spread function of a cylindrical lesion signal.
• Evaluated a ResNet-based DCNN model trained on patient images.

Main Results:

• Spatial resolution degradation in DCNN reconstructions worsened with decreased contrast, lower radiation dose, or increased DCNN denoising strength.
• The 50%/10% MTF spatial frequencies for the highest denoising strength varied from 0.15/0.30 mm⁻¹ to 0.36/0.72 mm⁻¹ across different contrasts.
• Filtered back projection (FBP) showed relatively constant 50%/10% MTF values of 0.38/0.76 mm⁻¹.

Conclusions:

• The proposed patient-data-based framework provides a more relevant assessment of DCNN spatial resolution in clinical CT.
• DCNN noise reduction significantly impacts spatial resolution, with the degree of impact dependent on image contrast, radiation dose, and denoising level.
• Careful selection of DCNN parameters is necessary to balance noise reduction and preservation of spatial resolution in CT imaging.