Multi-scale context UNet-like network with redesigned skip connections for medical image segmentation

Area of Science:

• Computer Vision
• Deep Learning
• Medical Image Analysis

Background:

• Medical image segmentation is crucial for intelligent medical assistant systems.
• UNet-like networks, including UNet2+ and UNet3+, have shown success but have limitations.
• UNet2+ struggles with comprehensive scale information, while UNet3+ has issues with small object segmentation.

Purpose of the Study:

• To propose a novel network topology, UNet_sharp (UNet#), for improved medical image segmentation.
• To enhance the integration of multi-scale feature maps for better understanding of organ/lesion positions and boundaries.
• To improve segmentation accuracy and reduce false positives.

Main Methods:

• Developed UNet_sharp (UNet#) by combining dense and full-scale skip connections in the decoder.
• Implemented deep supervision for model pruning to accelerate testing and enable mobile deployment.
• Incorporated two classification-guided modules to reduce false positives and enhance segmentation accuracy.

Main Results:

• UNet_sharp achieved superior Intersection over Union (IoU) and F1 scores compared to existing UNet-like networks.
• Demonstrated high performance in semantic segmentation tasks for nuclei, brain tumors, liver, and lung nodules.
• Achieved IoU values ranging from 74.01% to 95.36% and F1 scores from 84.77% to 97.34% across different tasks.

Conclusions:

• The reconstructed skip connections in UNet_sharp effectively integrate multi-scale contextual semantic information.
• UNet_sharp demonstrates superior accuracy in locating organs and lesions and segmenting boundaries compared to state-of-the-art models.
• The proposed method advances the development of accurate and efficient medical image segmentation for clinical applications.