Sparse Coding Inspired LSTM and Self-Attention Integration for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for clinical diagnosis and analysis.
• Integrating contextual relationships enhances neural network representational ability.
• Long Short-Term Memory (LSTM) and Self-Attention (SA) capture global dependencies but are typically separate modules.

Purpose of the Study:

• To present an innovative integration of LSTM design with SA sparse coding.
• To leverage LSTM's state compression and historical data retention for SA.
• To enhance SA's sparse coding and global dependency capture by incorporating temporal information.

Main Methods:

• Developed a novel approach using linear combinations of LSTM states for SA's Query, Key, and Value (QKV) matrices.
• Introduced two modules that integrate the SA matrix into the LSTM state design.
• Embedded these modules into a U-shaped convolutional neural network architecture for 2D and 3D medical images.

Main Results:

• The proposed modules significantly improve performance on medical image segmentation tasks across four datasets.
• Outperformed various baselines, including those already using contextual modules.
• Enhanced prediction accuracy in downstream segmentation tasks.

Conclusions:

• The integrated LSTM and SA sparse coding approach effectively models global dependencies.
• The novel modules enhance medical image segmentation accuracy without additional computational cost.
• This integration offers a promising direction for advancing medical image analysis.