Nonrigid registration method for longitudinal chest CT images in COVID-19

Area of Science:

• Medical imaging analysis
• Deep learning applications in radiology

Background:

• Analyzing morphological changes in COVID-19 pneumonia requires precise lung registration.
• Existing methods struggle with respiratory variations and lesion deformation.

Purpose of the Study:

• To develop and validate a nonrigid registration technique for COVID-19 pneumonia CT analysis.
• To ensure accurate tracking of lung changes without deforming lesion areas.

Main Methods:

• A deep learning-based nonrigid registration method (VoxelMorph) was employed for lung field alignment.
• Preprocessing involved classical registration for initial matching.
• Quantitative (image similarity, feature analysis) and qualitative (physician visual evaluation) assessments were performed.

Main Results:

• The method achieved high image similarity across 509 patient datasets.
• Analysis confirmed preservation of histogram characteristics and lesion features post-registration.
• Physician evaluation validated natural deformation of lung fields and lesions.

Conclusions:

• The developed deep learning nonrigid registration is effective for quantitative time-series analysis of lung changes in COVID-19 pneumonia.
• This technique offers a reliable tool for monitoring disease progression and treatment response.