Rib segmentation in chest x-ray images based on unsupervised domain adaptation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• 2D chest X-rays are widely used but challenging for accurate rib segmentation due to anatomical complexity and artifacts.
• Current deep learning methods require extensive pixel-level annotations, which are difficult to obtain for chest X-rays.
• Existing methods often struggle with fractured rib prediction and post-processing.

Purpose of the Study:

• To develop an efficient and accurate rib segmentation method for 2D chest X-ray images.
• To overcome the limitations of data annotation and post-processing in current rib segmentation techniques.
• To propose a cross-modal approach utilizing 3D CT data to guide 2D X-ray segmentation.

Main Methods:

• A novel cross-modal method based on unsupervised domain adaptation is proposed.
• A centerline loss function is incorporated to ensure result continuity and simplify post-processing.
• Digital reconstruction radiography images and 3D CT labels are used to train the model for segmenting unlabeled 2D chest X-ray images.

Main Results:

• The proposed model achieved a higher Dice score on test samples compared to existing methods.
• The segmentation results are highly interpretable and do not require manual rib markings on 2D X-rays.
• The method effectively addresses challenges of discontinuity and complex post-processing.

Conclusions:

• The developed unsupervised domain adaptation method offers an efficient and accurate solution for rib segmentation in 2D chest X-rays.
• This approach significantly reduces the need for manual annotation, a major bottleneck in medical image analysis.
• The method demonstrates potential for improved clinical applications requiring precise rib structure identification.