Unsupervised computed tomography and cone-beam computed tomography image registration using a dual attention network

Area of Science:

• Medical Imaging
• Radiotherapy
• Machine Learning

Background:

• Accurate registration of CT and CBCT images is crucial for image-guided radiotherapy (IGRT).
• Significant intensity variations between CT and CBCT images pose challenges to traditional registration methods.
• Existing methods often struggle with performance and clinical applicability due to these intensity differences.

Purpose of the Study:

• To develop a learning-based unsupervised approach for accurate CT and CBCT image registration.
• To address the limitations posed by large intensity variations in IGRT applications.
• To predict the deformation field for precise image alignment.

Main Methods:

• A dual attention module, incorporating scale-aware position (SP-BLOCK) and channel attention (SC-BLOCK) blocks, was employed.
• The SP-BLOCK integrates multi-scale features to enhance feature correlation across different positions.
• The SC-BLOCK selectively emphasizes channel dependencies to manage diverse feature information.

Main Results:

• The proposed method achieved superior performance on the 4D-LUNG dataset compared to existing techniques.
• Highest Structural Similarity Index (SSIM) of 86.34% and Dice Similarity Coefficient (DICE) of 89.74% were recorded.
• The lowest Target Registration Error (TRE) of 2.07 mm demonstrated high accuracy.

Conclusions:

• The developed unsupervised method achieves high-accuracy CT-CBCT image registration without manual labeling.
• This technique offers an effective solution for precise patient positioning and target localization in IGRT.
• The findings support the clinical utility of the proposed method in enhancing radiotherapy precision.