2D medical image segmentation via learning multi-scale contextual dependencies.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Deep Learning for Segmentation

Background:

• Automatic medical image segmentation is crucial for disease diagnosis and treatment planning.
• Convolutional Neural Networks (CNNs) show promise but struggle with small regions of interest (ROIs) and limited context, leading to high missed diagnosis rates.
• Existing models often overlook the challenges posed by small ROIs and insufficient contextual information in medical imaging.

Purpose of the Study:

• To develop a new segmentation framework that improves the representation of small ROIs, especially in deeper network layers.
• To explicitly learn global contextual dependencies across multi-scale feature spaces for enhanced segmentation accuracy.
• To boost the interpretability of neural networks through feature visualization.

Main Methods:

• Proposed a novel segmentation framework designed to enhance the representative capability of small ROIs.
• Implemented adaptive aggregation of local features and global dependencies across spatial and channel dimensions.
• Integrated visualization techniques to improve the interpretability of learned features.

Main Results:

• Achieved state-of-the-art performance in liver tumor segmentation (91.18% Sensitivity).
• Demonstrated high accuracy in COVID-19 lung infection segmentation (75.73% Sensitivity).
• Showcased effectiveness in retinal vessel detection (82.68% Sensitivity).

Conclusions:

• The proposed framework significantly reduces missed diagnoses in medical image segmentation tasks.
• The method effectively enhances small ROI representation and captures global contextual dependencies.
• The framework is adaptable and can be integrated into existing Fully Convolutional Network (FCN) models to improve their performance.