CNN Segmentation via Multi-Scale Context Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CNN models for medical image segmentation, such as U-Net, face challenges in achieving high segmentation accuracy while maintaining or reducing computation requirements, especially in resource-constrained medical devices and datasets with limited training images.
Innovation Solution
The proposed CNN model incorporates a multi-scale context aggregation module and a Channel-wise Feature Selection (CFS) module to enhance segmentation accuracy and reduce computation requirements. The multi-scale context aggregation module aggregates contextual information from global to local scales using atrous-convolution feature maps, while the CFS module merges feature maps efficiently using attention mechanisms and element-wise operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If lightweight networks (ENet, ESPNet, Fast-SCNN) are used to reduce computation requirements, then computation requirements are reduced, but segmentation accuracy becomes insufficient for most medical applications
Solution Approach 1:
The network is segmented into distinct functional modules: an encoder for feature extraction, a multi-scale context aggregation module for contextual information integration, and a decoder for segmentation map generation. This modular segmentation allows each component to be optimized independently, achieving high accuracy while controlling computation requirements.
Solution Approach 2:
The patent introduces multi-scale context aggregation that processes features at multiple spatial scales simultaneously. By aggregating contextual information from different scales (from fine-grained to coarse-grained), the network achieves improved segmentation accuracy without proportionally increasing computation requirements, as the multi-scale processing is integrated efficiently into the existing network architecture.
2Measurement precision
If extensions and modifications to U-Net are made to improve segmentation accuracy, then segmentation accuracy is improved, but computation cost increases significantly (U-Net++ requires 2.5 times more computation than U-Net)
Solution Approach 1:
The patent extracts and integrates only the essential multi-scale context aggregation functionality into the U-Net architecture, rather than adopting the entire complex structure of U-Net++. By selectively incorporating multi-scale processing at strategic points in the network, the patent achieves accuracy improvements without incurring the full computation cost of U-Net++.
Solution Approach 2:
Instead of implementing full multi-scale processing throughout the entire network (which would be computationally expensive), the patent applies multi-scale context aggregation selectively at key stages where contextual information is most beneficial. This partial application of multi-scale processing achieves most of the accuracy benefits while avoiding excessive computation costs.
3Measurement precision
If large CNN models (PSPNet with 66M parameters, DeepLabv3+ with 41M parameters) are deployed to achieve high segmentation accuracy, then segmentation accuracy is improved, but deployment on cutting-edge medical devices becomes difficult due to optimization bias toward medical functionalities rather than computing power
Solution Approach 1:
The patent employs parameter sharing and weight initialization strategies that reduce the total number of parameters while maintaining representational capacity. By carefully designing the network to share parameters across different scales and stages, the patent achieves high segmentation accuracy with a significantly reduced parameter count suitable for medical device deployment.
Solution Approach 2:
The network is designed with universal feature extraction capabilities that can handle multiple medical imaging modalities and segmentation tasks. The multi-scale context aggregation module serves multiple functions: it captures contextual information, handles varying object sizes, and adapts to different imaging modalities, reducing the need for modality-specific or task-specific network variants.
4Reliability
If large CNN models are trained on medical datasets with limited training images, then training becomes difficult, but using smaller networks reduces segmentation accuracy
Solution Approach 1:
The network incorporates pre-trained weights from large-scale natural image datasets before fine-tuning on medical images. This preliminary training on abundant non-medical data initializes the network with general feature extraction capabilities, enabling effective training on limited medical datasets while achieving high segmentation accuracy.
Solution Approach 2:
The patent employs data augmentation techniques that continuously generate diverse training samples from limited medical images through transformations such as rotations, flips, and elastic deformations. This continuous generation of augmented training data maintains training feasibility while preserving the network's ability to learn accurate segmentation patterns.
Data Source
AI summary
In a convolutional neural network (CNN) using an encoder-decoder structure for image segmentation, a multi-scale context aggregation module receives an encoded final-stage feature map from the encoder, and sequentially aggregates multi-scale contexts of this feature map from a global scale to a local scale to strengthen semantic relationships of contexts of different scales to improve segmentation accuracy. The multi-scale contexts are obtained by computing atrous convolution on the feature map for different dilation rates. To reduce computation, a channel-wise feature selection (CFS) module is used in the decoder to merge two input feature maps. Each feature map is processed by a global pooling layer followed by a fully connected layer or a 1×1 convolutional layer to select channels of high activation. By subsequent channel-wise multiplication and elementwise summation, only channels with high activation in both feature maps are preserved and enhanced in the merged feature map.