Global-Local Feature Reconstruction for Medical Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image segmentation methods, such as U-Net, suffer from insufficient global feature extraction, long-distance feature dependency modeling, and insufficient recovery of spatial information, leading to inaccurate segmentation results.
Innovation Solution
A global and local feature reconstruction network is introduced, incorporating a global feature reconstruction module (GFR) to transfer semantic information from high-level to low-level features and a local feature reconstruction module (LFR) for dynamic upsampling, enhancing feature fusion and spatial information recovery.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If simple skip connections are used to fuse features at different levels, then the network structure remains simple, but global context extraction capability is insufficient
Solution Approach 1:
The patent introduces a global feature reconstruction module as an intermediary component that receives high-level semantic features and reconstructs them to low-level spatial resolutions, mediating between encoder and decoder features to enhance global context extraction without significantly complicating the overall network structure
Solution Approach 2:
The feature fusion process is segmented into distinct functional modules: global feature reconstruction module for capturing global context, local feature reconstruction module for preserving spatial details, and feature fusion module for combining them, allowing each module to specialize in specific aspects of feature processing
2Measurement precision
If non-local module is used to model long-range features dynamically, then pixel-level feature representation is enhanced, but computational complexity increases significantly
Solution Approach 1:
The attention mechanism is segmented into channel attention (SE module) and spatial attention modules, processing features in stages rather than computing all pixel pairs simultaneously, reducing peak computational complexity while maintaining pixel-level representation quality
Solution Approach 2:
The patent uses partial attention mechanisms that focus computational resources on relevant feature channels and spatial locations rather than computing attention for all feature pairs, achieving sufficient pixel-level representation with reduced computational burden
3Loss of information
If SE module is used to extract global spatial features, then channel features are enhanced, but spatial awareness for pixel-level segmentation is insufficient
Solution Approach 1:
The attention mechanism is divided into separate channel attention (SE module) and spatial attention modules, where channel attention enhances feature representations and spatial attention restores spatial awareness, with each module performing its specialized function in sequence
Solution Approach 2:
The network maintains continuous feature flow through skip connections that preserve spatial information from encoder to decoder, ensuring that spatial awareness is continuously maintained throughout the network rather than being lost and recovered
4Ease of manufacture
If nearest neighbor or bilinear interpolation is used for feature upsampling, then implementation is simple, but spatial information recovery is insufficient
Solution Approach 1:
Traditional mechanical interpolation methods (nearest neighbor, bilinear) are replaced with a learnable local feature reconstruction module that uses convolutional operations to dynamically reconstruct spatial features, achieving superior spatial information recovery while maintaining implementation simplicity through modular design
5Manufacturing precision
If transposed convolution is used for upsampling, then spatial information is recovered, but semantic information at different locations is ignored
Solution Approach 1:
The upsampling process is segmented into semantic feature processing by the global feature reconstruction module followed by spatial detail recovery by the local feature reconstruction module, allowing each module to optimize for its specific function without compromise
Data Source
AI summary
The invention provides a global and local feature reconstruction network-based medical image segmentation method, device, apparatus and a computer storage medium. In the invention, a global feature reconstruction GFR module introduces semantic information from high-level into low-level features through a global descriptor, to eliminate a semantic gap between features in different levels, and a feature map is reconstructed by using a cross-level global descriptor, to implement long-distance feature dependency modeling. A local feature reconstruction LFR module implements dynamic upsampling of features based on the guidance of a low-stage feature map. Local features are dynamically reconstructed, so that spatial detail information in the low-stage feature map is transferred to a high-stage feature map, thereby implementing spatial information recovery of features, overcoming the problems such as insufficient global feature extraction and long-distance feature dependency modeling capabilities and insufficient recovery of spatial information of a feature map in a U-shaped encoder-decoder network.