MDU-Net: multi-scale densely connected U-Net for biomedical image segmentation.

Area of Science:

• Biomedical image analysis
• Deep learning for medical imaging

Background:

• Biomedical image segmentation is crucial for medical analysis and diagnosis.
• Deep convolutional networks (DNNs), particularly U-Net, have advanced segmentation.
• Existing methods face challenges in feature propagation and network depth.

Purpose of the Study:

• To propose a Multi-scale Densely connected U-Net (MDU-Net) for improved biomedical image segmentation.
• To enhance feature propagation and enable deeper U-Net architectures.
• To investigate the impact of quantization on segmentation performance.

Main Methods:

• Developed three Multi-scale Dense Connections (MDC) for U-Net's encoder, decoder, and across them.
• Integrated MDCs into a novel MDU-Net architecture.
• Introduced quantization to mitigate overfitting in dense connections.

Main Results:

• The proposed MDCs improved U-Net performance by up to 1.8% (Test A) and 3.5% (Test B) on the MICCAI 2015 Gland Segmentation dataset.
• MDU-Net with quantization demonstrated significant improvements over the original U-Net.
• Dense connections facilitated the creation of deeper U-Net models.

Conclusions:

• The MDU-Net architecture effectively enhances biomedical image segmentation.
• Multi-scale dense connections are a promising approach for improving U-Net performance.
• Quantization further refines segmentation accuracy and model robustness.