Adaptation to CT Reconstruction Kernels by Enforcing Cross-Domain Feature Maps Consistency

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning aids COVID-19 analysis in chest CT images.
• Domain shift, particularly from reconstruction kernels, degrades model performance.
• Robust algorithms are needed for clinical application with larger datasets.

Purpose of the Study:

• To address the domain shift problem in COVID-19 CT image segmentation.
• To compare existing domain adaptation techniques.
• To propose and validate a novel unsupervised domain adaptation method.

Main Methods:

• Investigated the impact of reconstruction kernel differences on COVID-19 segmentation.
• Compared task-specific augmentation and unsupervised adversarial learning.
• Proposed F-Consistency, an unsupervised method minimizing feature map MSE between paired CT images.

Main Results:

• F-Consistency achieved a 0.64 Dice Score on unseen sharp kernels, outperforming the baseline (0.56).
• F-Consistency improved paired image prediction Dice Score to 0.80 (baseline 0.46).
• The method demonstrated better generalization on unseen kernels and without lesions.

Conclusions:

• F-Consistency effectively mitigates domain shift caused by varying CT reconstruction kernels.
• The proposed method offers a robust solution for clinical deployment of deep learning in COVID-19 CT analysis.
• Unsupervised adaptation using paired images with differing kernels is a promising direction.