Quaternion Cross-Modality Spatial Learning for Multi-Modal Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep Neural Networks (DNNs) are impactful in medical image segmentation.
• Real-valued convolution in DNNs is used for multi-modal segmentation but struggles with spatial dependence.
• Maintaining spatial dependence is critical for accurate lesion distribution identification.

Purpose of the Study:

• To propose a novel method, Quaternion Cross-modality Spatial Learning (Q-CSL), for improved multi-modal medical image segmentation.
• To address the limitations of weighted summation in conventional convolutions regarding spatial information.
• To enhance the learning and fusion of spatial information across different imaging modalities.

Main Methods:

• Introduction of quaternion representation for data and coordinates to capture spatial information.
• Development of Quaternion Spatial-association Convolution for learning spatial features.
• Proposal of the De-level Quaternion Cross-modality Fusion (De-QCF) module for feature excavation and cross-modality spatial dependency fusion.

Main Results:

• The proposed Q-CSL method demonstrates strong performance compared to existing methods.
• The approach achieves high accuracy in multi-modal medical image segmentation.
• The method is computationally efficient, utilizing only 0.01061 M parameters and 9.95G FLOPs.

Conclusions:

• Q-CSL effectively explores spatial information and cross-modality linkages in medical images.
• The novel quaternion-based approach overcomes limitations of traditional convolutions for segmentation.
• The method offers a promising, efficient solution for complex medical image analysis tasks.