Kernel Conversion for Robust Quantitative Measurements of Archived Chest Computed Tomography Using Deep...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Quantitative Imaging

Background:

• Chest computed tomography (CT) is crucial for lung cancer screening and evaluating pulmonary/extra-pulmonary abnormalities.
• Sharp-kernel images, used for visual assessment, have increased noise, while soft-kernel images are needed for accurate quantification.
• Archived sharp-kernel images are often unusable for quantitative analysis due to storage limitations.

Purpose of the Study:

• To develop deep neural network (DNN) models for converting sharp-kernel CT images to soft-kernel-like images.
• To enable the reuse of historical chest CT data for robust quantitative measurements.
• To validate the accuracy and clinical applicability of the DNN-based kernel conversion method.

Main Methods:

• Trained DNN models using paired sharp-kernel (input) and soft-kernel (ground-truth) images from 30 COPD patients.
• Evaluated model accuracy on independent CT scans from 30 smokers with/without COPD.
• Assessed conversion error against phantom scan variability and compared Dice coefficients with other filtering methods.

Main Results:

• DNN-converted images showed CT value differences comparable to the CT device's measurement error.
• Dice coefficients for low attenuation voxels were significantly higher for DNN-converted images (p < 0.001).
• Good agreement was found in quantitative measurements of emphysema, intramuscular adipose tissue, and coronary artery calcification.

Conclusions:

• The developed DNN model effectively converts sharp-kernel CT images to soft-kernel-like images.
• This method enables robust quantitative analysis of archived chest CT scans, enhancing longitudinal study value.
• The phantom-based validation approach is applicable to other deep learning image conversion techniques.