A new method to validate thoracic CT-CT deformable image registration using auto-segmented 3D anatomical landmarks

Area of Science:

• Medical Imaging
• Radiology
• Computational Anatomy

Background:

• Deformable image registration is crucial but prone to errors, especially in thoracic CT due to low lung contrast.
• Accurate identification of registration inaccuracies is vital for reliable medical analysis.
• Validating thoracic CT-CT registration using anatomical landmarks is an unmet need.

Purpose of the Study:

• To validate thoracic CT-CT image registration using auto-segmented anatomical landmarks.
• To assess the accuracy of different image registration algorithms in the thoracic region.
• To quantify registration inaccuracies using bronchial branch points.

Main Methods:

• Five lymphoma patients underwent serial CT scans over 18 months.
• Three registration algorithms (Demons, B-spline, Affine) were used to register follow-up CTs to a reference scan.
• Auto-segmented bronchial branch points and Dice Similarity Coefficients (DSC) were employed for evaluation.

Main Results:

• Median deviations for Demons, B-spline, and Affine were 1.6, 1.1, and 4.2 mm, respectively.
• Maximum deviations exceeding 15 mm were observed with Demons and B-spline algorithms.
• DSC values indicated similar discrepancies across algorithms (0.96, 0.97, 0.91).

Conclusions:

• Auto-segmented bronchial branch points are effective for validating thoracic CT-CT image registration.
• This method successfully identified local registration errors greater than 15 mm.
• The findings highlight the utility of anatomical landmarks for assessing registration accuracy in challenging thoracic datasets.