Dense gate network for biomedical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Biomedical image semantic segmentation is crucial for medical diagnosis.
• Existing encoder-decoder models with skip connections struggle with semantic gaps between fused features.
• These gaps hinder the overall effectiveness of deep learning networks in this domain.

Purpose of the Study:

• To introduce a novel Dense Gate Network (DG-Net) for improved biomedical image semantic segmentation.
• To address the challenge of semantic gaps in encoder-decoder architectures.
• To enhance feature fusion by aligning semantic levels between encoder and decoder features.

Main Methods:

• Proposed Dense Gate Network (DG-Net) utilizing a Gate Aggregate structure.
• Employed a gate unit to minimize categorical ambiguity and guide feature recovery.
• Features from encoder and decoder are processed to reduce semantic gaps before fusion.

Main Results:

• DG-Net achieved high Intersection over Union (IOU) scores across four datasets (CT and microscopy images).
• Achieved IOU scores of 97.953%, 89.569%, 81.870%, and 76.486% in cross-validation experiments.
• Outperformed U-Net and MultiResUNet in average IOU and accuracy.

Conclusions:

• DG-Net demonstrates competitive performance against baseline methods in biomedical image segmentation.
• The Gate Aggregate structure and gate unit effectively improve network performance by feature aggregation and reducing semantic gaps.
• DG-Net shows significant potential for advancing biomedical image segmentation applications.