CvTMorph: Improving Local Feature Extraction in Medical Image Registration for Respiratory Motion Modeling with...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate respiratory motion modeling in 4D medical images is vital for applications like radiation therapy planning.
• Current registration methods often fail to effectively capture local features, limiting their performance.

Purpose of the Study:

• To introduce CvTMorph, a novel framework designed to enhance local feature extraction for improved respiratory motion modeling.
• To leverage the strengths of Convolutional vision Transformers (CvT) and Convolutional Neural Networks (CNN) in a hybrid approach.

Main Methods:

• CvTMorph integrates CvT and CNN into a hybrid model, enhanced with scaling and square layers for superior registration.
• Performance was evaluated on the 4D-Lung and DIR-Lab datasets, with comparisons against state-of-the-art methods.

Main Results:

• CvTMorph demonstrated superior accuracy and robustness in respiratory motion modeling compared to existing methods.
• The integration of CvT significantly boosted registration performance and the representation of local image structures.

Conclusions:

• CvTMorph presents a potent solution for precise respiratory motion modeling in 4D medical imaging.
• The hybrid CvT-CNN model effectively extracts local features, enhancing registration and showing promise for radiation therapy planning.