Multi-Scale Self-Guided Attention for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Standard convolutional neural networks (CNNs) for medical image segmentation face challenges with redundant feature extraction and inefficient modeling of long-range dependencies.
• Existing encoder-decoder architectures often extract similar low-level features multiple times, and struggle to capture global context effectively.

Purpose of the Study:

• To address limitations in current medical image segmentation models by proposing a novel architecture.
• To enhance the capture of contextual dependencies and improve feature representation using guided self-attention mechanisms.

Main Methods:

• Developed a new architecture incorporating guided self-attention mechanisms to integrate local features with global dependencies.
• Employed an additional loss function to guide attention, focusing on relevant image regions and feature associations.
• Evaluated the model on diverse medical imaging datasets, including abdominal organs, cardiovascular structures, and brain tumors.

Main Results:

• The proposed model demonstrated superior segmentation performance compared to state-of-the-art methods.
• Achieved increased prediction accuracy and reduced standard deviation in segmentation results.
• Ablation experiments confirmed the significant contribution of the attention modules to the model's effectiveness.

Conclusions:

• The novel guided self-attention approach effectively overcomes limitations of standard CNNs in medical image segmentation.
• The model generates precise and reliable automatic segmentations by adaptively highlighting interdependent channel maps and focusing on discriminant regions.
• The publicly available code facilitates further research and application of this advanced segmentation technique.