Stagger Network: Rethinking information loss in medical image segmentation with various-sized targets.

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning for medical imaging

Background:

• Medical image segmentation requires models to capture both local and global information, especially for targets of varying sizes.
• Current CNN and ViT approaches struggle to balance multi-scale target detection, often leading to significant information loss due to divergent feature distributions.

Purpose of the Study:

• To introduce a novel Stagger Network (SNet) designed to mitigate information loss in medical image segmentation.
• To develop a fusion structure that effectively balances local and global feature extraction from CNNs and ViTs for improved multi-scale segmentation.

Main Methods:

• Proposed a novel Stagger Network (SNet) incorporating a Parallel Module to bridge semantic gaps between CNN and ViT features.
• Introduced a Stagger Module for fusing semantically similar features and an Information Recovery Module to recapture complementary information.
• Theoretically analyzed the parallel and stagger strategies to demonstrate reduced information loss.

Main Results:

• The proposed SNet demonstrated superior performance in segmenting various-sized targets on the Synapse dataset compared to state-of-the-art methods.
• SNet also showed superiority on the ACDC and MoNuSeg datasets, which feature targets with more consistent dimensions.

Conclusions:

• The SNet effectively addresses the challenge of multi-scale medical image segmentation by intelligently fusing CNN and ViT features.
• The proposed fusion strategies significantly reduce information loss, leading to enhanced segmentation accuracy across datasets with diverse target size variations.