PS5-Net: a medical image segmentation network with multiscale resolution.

Area of Science:

• Artificial Intelligence
• Medical Image Analysis
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) are integral to clinical diagnostic support, especially in medical image segmentation.
• U-Net architectures with skip connections are widely used but face challenges with complex organs, morphological variations, and single-resolution limitations.
• Existing methods struggle with information loss due to pooling and limited contextual interaction in U-Net structures.

Purpose of the Study:

• To address limitations in current deep neural networks for medical image segmentation.
• To develop a multiscale segmentation network capable of handling complex anatomical structures and variations.
• To improve feature extraction and integration across different resolutions for enhanced segmentation accuracy.

Main Methods:

• Proposed the five-layer pyramid segmentation network (PS5-Net), a multiscale network based on the U-Net architecture.
• Implemented a kernel selection module to weight and fuse features from diverse resolutions.
• Enhanced the U-Net feature extraction network (PS-UNet) with dilated convolutions while preserving the classical structure.

Main Results:

• Achieved a Dice score of 0.9613 for liver segmentation on the CHLISC dataset.
• Attained a Dice score of 0.8587 for skin lesion segmentation on the ISIC2018 dataset.
• Demonstrated superior performance compared to existing medical image segmentation methodologies.

Conclusions:

• PS5-Net effectively utilizes multiscale semantic information and global contextual connections for accurate segmentation.
• The network shows superior performance in complex medical image segmentation tasks, offering potential for clinical applications.
• PS5-Net advancements contribute to improved diagnostic and analytical processes, enhancing patient care through precise image analysis.