Evaluation of various deformable image registration algorithms for thoracic images

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Computational Anatomy

Background:

• Accurate deformable image registration (DIR) is crucial for thoracic 4D CT applications.
• Evaluating DIR algorithm performance in oncology is essential for treatment planning and analysis.

Purpose of the Study:

• To assess the spatial accuracy of four deformable image registration algorithms for thoracic 4D CT.
• To compare a commercial algorithm (B-spline in Velocity AI) against three public domain algorithms (FFD, Horn-Schunk optical flow, Demons).

Main Methods:

• Five esophageal cancer patient datasets with manually identified anatomical landmarks were used.
• Manually measured displacement vector fields (mDVF) were compared against algorithm-calculated displacement vector fields (aDVF).
• Registration error was quantified as the difference between mDVF and aDVF, with mean 3D errors calculated for each algorithm.

Main Results:

• Mean 3D registration errors were: B-spline (2.7 ± 0.8 mm), FFD (3.6 ± 1.0 mm), Optical Flow (2.4 ± 0.9 mm), and Demons (2.4 ± 1.2 mm).
• B-spline, Optical Flow, and Demons algorithms demonstrated reasonable accuracy.
• The B-spline algorithm in Velocity AI showed small errors for displacements under ~10 mm.

Conclusions:

• B-spline, Optical Flow, and Demons algorithms show potential for 4D dose calculation, image segmentation, and ventilation imaging in thoracic cancer.
• Optimizing algorithm parameters may further enhance registration accuracy.
• The Velocity AI B-spline algorithm is potentially useful for thoracic 4D CT applications with displacements up to ~10 mm.