Source-linked AI summary
Channel prior convolutional attention for medical image segmentation
Hejun Huang, Zuguo Chen, Ying Zou, Ming Lu, Chaoyang Chen
TL;DR
Medical image segmentation is challenged by low-contrast images, organ-shape variation, and attention mechanisms with insufficient adaptive capability. The paper proposes CPCA, which dynamically distributes channel and spatial attention using multi-scale depth-wise convolutions, and builds CPCANet around it. CPCANet achieves enhanced segmentation performance with fewer computational resources, while accurate boundaries and variable network sizes remain limitations.
Problem
Existing attention mechanisms have insufficient adaptive capability for medical image segmentation, particularly in dynamically distributing attention across channels and spatial regions.
Method
CPCA combines channel attention with multi-scale depth-wise convolutional spatial attention computed independently for each channel, and replaces Transformer self-attention in CPCANet.
Results
CPCANet achieves enhanced segmentation performance while reducing computational complexity compared with existing methods and is validated on two publicly available datasets.
Takeaways & Limitations
CPCA enables attention to focus dynamically on informative channels and important regions in medical image segmentation.
Takeaways & Limitations
CPCA has difficulty achieving accurate segmentation boundaries, and CPCANet is limited to a single network size and two validated datasets.
Abstract
from arXiv · showhide
Characteristics such as low contrast and significant organ shape variations are often exhibited in medical images. The improvement of segmentation performance in medical imaging is limited by the generally insufficient adaptive capabilities of existing attention mechanisms. An efficient Channel Prior Convolutional Attention (CPCA) method is proposed in this paper, supporting the dynamic distribution of attention weights in both channel and spatial dimensions. Spatial relationships are effectively extracted while preserving the channel prior by employing a multi-scale depth-wise convolutional module. The ability to focus on informative channels and important regions is possessed by CPCA. A segmentation network called CPCANet for medical image segmentation is proposed based on CPCA. CPCANet is validated on two publicly available datasets. Improved segmentation performance is achieved by CPCANet while requiring fewer computational resources through comparisons with state-of-the-art algorithms. Our code is publicly available at \url{https://github.com/Cuthbert-Huang/CPCANet}.
1. Introduction
Medical image segmentation needs adaptive feature selection because CNNs have limited receptive fields and Transformers generalize insufficiently on small datasets. The paper introduces CPCA and CPCANet to distribute attention dynamically across channels and spatial regions while reducing computational complexity.
- CNNs struggle to capture global information because of limited receptive fields, while Transformer generalization remains insufficient on small datasets.
- CPCA dynamically distributes attention weights across both channel and spatial dimensions.
- Multi-scale depth-wise strip convolutions extract spatial relationships while reducing computational complexity and preserving channel-specific attention.
- CPCANet integrates CPCA into a Vision Transformer backbone, replacing self-attention to focus on valuable channels and significant regions.
- The proposed network improves segmentation performance while reducing computational complexity compared with existing methods.
2. Related work
Related work combines CNNs, Transformers, and attention mechanisms to address local-context limitations and long-range dependencies in medical image segmentation. CPCA extends channel-spatial attention by computing spatial attention independently for each channel with multi-scale depth-wise convolutions.
- U-Net variants commonly incorporate attention, dense connections, skip connections, and deep supervision for medical image segmentation.
- Vision Transformers model long-range dependencies but often lack strong local inductive bias and require large datasets, motivating CNN-Transformer hybrids.
- Attention mechanisms separately emphasize important objects through channels and significant regions through spatial features.
- CBAM combines channel and spatial attention, but channel compression produces a consistent spatial attention distribution across channels.
- CPCA computes a separate spatial attention map for each channel using multi-scale depth-wise stripe convolutional kernels.
3. Methods
The method sequentially applies channel and spatial attention, using channel priors and multi-scale depth-wise convolutions to produce dynamically distributed attention maps with reduced computational complexity.
- Design objectives: CPCA targets dynamic attention-weight distribution across both channel and spatial dimensions while avoiding the computational overhead of complex self-attention.Uniform spatial weights across channels can introduce noise, while self-attention becomes burdensome for large-resolution and 3D images.
- Sequential attention: The channel attention map is multiplied with the input feature to form a channel prior that is then processed by spatial attention.For input F ∈ R^C×H×W, channel attention produces Mc ∈ R^C×1×1 and refined features Fc ∈ R^C×H×W.
- Channel attention: Channel attention aggregates spatial information with average and max pooling, processes both descriptors through a shared MLP, and combines them into a channel attention map.The shared MLP uses one hidden layer with size C/r×1×1, where r is the reduction ratio.
- Spatial attention: Spatial attention uses multi-scale depth-wise convolutions to extract pixel relationships while preserving inter-channel relationships and reducing computational complexity.Different channel-wise kernel sizes and depth-wise stripe convolutions capture multi-scale information; a 1×1 convolution performs channel mixing.
- Attention arrangement: Channel and spatial attention are arranged sequentially because experimental results reported in the paper favor this arrangement over parallel placement.The spatial module operates on the preceding channel-attended representation, forming channel prior convolutional attention.
- CPCANet architecture: CPCANet adopts a pyramid encoder whose CPCA blocks replace self-attention, while decoder experiments select Conv blocks over CPCA blocks.Four encoder stages with decreasing resolutions correspond to four decoder stages with increasing resolutions; convolution and de-convolution stems accommodate high-resolution inputs.
4. Experiments and results
CPCANet was evaluated on automated cardiac diagnosis and skin lesion segmentation using public datasets, with quantitative and visual comparisons against existing methods. It achieved strong segmentation results while requiring reduced computational resources in cardiac diagnosis.
- Datasets: The ACDC dataset contains 100 samples and segments the right ventricle, myocardium, and left ventricle using DSC and HD95 metrics.After preprocessing, 1,902 slices were divided into training, validation, and testing sets.
- Datasets: The skin lesion experiments used ISIC-2016 for training and PH2 for testing, evaluating DSC and IoU.ISIC-2016 contains 900 lesion-segmentation samples.
- Automated cardiac diagnosis: CPCANet achieved an average DSC of 92.60% with 10.62G FLOPs in automated cardiac diagnosis.These results were obtained from a single model without ensembles, pre-training, or additional data.
- Automated cardiac diagnosis: CPCANet achieved an HD95 of 1.097mm in cardiac diagnosis, representing a 10.4% absolute error reduction compared with nnUNet.The comparison used HD95 alongside DSC and computational requirements.
- Visual comparisons: Visual comparisons show CPCANet concentrating on important objects and regions, with improved details in cardiac and challenging skin-lesion areas.The reported challenging regions include fine branches and local protrusions, while cardiac comparisons note missegmentation by nnUNet.
- Skin lesion segmentation: CPCANet achieved a DSC of 93.7% and an IoU of 88.8% for skin lesion segmentation, improving over BAT by 1.6 and 3 percentage points, respectively.The experiments compared single-model results without ensembles, pre-training, or additional data.
5. Discussion
Ablation and component comparisons support combining channel and spatial attention sequentially, using multi-scale convolutional kernels and channel mixing. The method still has difficulty producing accurate segmentation boundaries and is limited to a single network size.
- Ablation study: Combining channel and spatial attention produces better segmentation performance than using either attention type alone.The ablation evaluates channel-only, spatial-only, parallel, and sequential configurations on ACDC.
- Ablation study: Sequential channel-spatial attention performs better than placing the two attention modules in parallel.This ordering supports the proposed sequential design rationale.
- Spatial attention kernel size: Larger convolutional kernels capture global information, whereas smaller kernels are effective for texture information.Multi-scale kernels therefore address both semantic context and pixel-level detail requirements.
- Effects of channel mixing: CPCA without channel mixing has similar DSC to CBAM but a noticeable HD95 decrease, while channel mixing significantly improves DSC.The comparison indicates that channel mixing further enhances feature representation and CPCA’s region-focused behavior.
- Decoder design: ConvBlock achieves good segmentation performance with fewer building blocks and is selected for the decoder.The decoder uses three stages with 2, 2, and 1 building blocks, respectively.
- Limitations and future work: CPCA remains limited in accurately segmenting boundaries, and CPCANet is restricted to a single network size.Future work targets boundary accuracy, multiple network sizes, and validation on more diverse datasets.
6. Conclusion
The paper proposes CPCA and a CPCA-based segmentation network for medical image segmentation. CPCA dynamically distributes attention across channel and spatial dimensions while using depth-wise convolutions and channel mixing to enhance spatial representation.
- Conclusion: CPCA dynamically distributes attention weights across channel and spatial dimensions and supports multi-scale spatial relationship extraction.Its depth-wise convolution preserves the channel prior, while channel mixing further enhances feature representation.
- Conclusion: CPCANet achieves enhanced segmentation performance with reduced computational resource requirements on two publicly available datasets.The comparison is made against existing methods.