Source-linked AI summary
Attention Guided Network for Retinal Image Segmentation
Shihao Zhang, Huazhu Fu, Yuguang Yan, Yubing Zhang, Qingyao Wu, Ming Yang, Mingkui Tan, Yanwu Xu
TL;DR
Retinal CNN segmentation can lose structural information such as edges during convolution and pooling, despite the clinical importance of accurate vessel and optic disc/cup segmentation. AG-Net uses an attention guided filter as a structure-sensitive expanding path to transfer low-level structure, suppress background noise, and fuse multiresolution features. Experiments on vessel and optic disc/cup segmentation tasks demonstrate its effectiveness, including improvements over GF-Net on DRIVE and superiority to compared methods on ORIGA.
Problem
CNN convolutions and pooling may neglect retinal edge structures that are important for accurate vessel, optic disc, and cup segmentation.
Method
AG-Net integrates an attention guided filter into an M-Net-based network to replace skip connections and upsampling, transfer structural information, and reduce background noise.
Results
AG-Net outperforms GF-Net on DRIVE by 0.0010 Acc, 0.0019 AUC, 0.0205 Sen, and 0.0126 IOU, and outperforms compared OD/OC methods on ORIGA.
Takeaways & Limitations
Across vessel and optic disc/cup segmentation tasks, the experiments demonstrate the effectiveness of combining structural filtering with attention in the expanding path.
Abstract
from arXiv · showhide
Learning structural information is critical for producing an ideal result in retinal image segmentation. Recently, convolutional neural networks have shown a powerful ability to extract effective representations. However, convolutional and pooling operations filter out some useful structural information. In this paper, we propose an Attention Guided Network (AG-Net) to preserve the structural information and guide the expanding operation. In our AG-Net, the guided filter is exploited as a structure sensitive expanding path to transfer structural information from previous feature maps, and an attention block is introduced to exclude the noise and reduce the negative influence of background further. The extensive experiments on two retinal image segmentation tasks (i.e., blood vessel segmentation, optic disc and cup segmentation) demonstrate the effectiveness of our proposed method.
1 South China University of Technology, Guangzhou, China
The listed affiliations include institutions in Abu Dhabi, Shenzhen, Guangzhou, and Ningbo.
- The affiliations name the Inception Institute of Artificial Intelligence in Abu Dhabi and Peng Cheng Laboratory in Shenzhen.
- They also identify CVTE Research in Guangzhou and the Cixi Institute of Biomedical Engineering in Ningbo.
- The project page is hosted at github.com/HzFu/AGNet.
1 Introduction
Retinal segmentation requires preserving contextual structures such as vessels, optic discs, and cups, but CNN convolutions and pooling may neglect useful edges. AG-Net addresses this by using an attention guided filter as a structure-sensitive expanding path that transfers low-level structure while reducing background noise.
- Retinal vessels, optic discs, and cups provide clinical information, making accurate segmentation important for automatic disease diagnosis.
- Convolutions and pooling can neglect edge structures that are important for retinal image analysis.
- FCN and U-Net use skip connections but do not effectively leverage structural information, motivating a better expanding path.
- AG-Net incorporates guided filtering into CNNs to transfer low-level structural information to high-level features during expansion.
- An attention mechanism removes noisy components introduced by complex backgrounds while preserving structural information for vessel and optic disc/cup segmentation.
2 Methodology
AG-Net uses an attention guided filter within an M-Net-based expanding path to fuse multiresolution features while preserving spatial and structural information. The filter computes attention-weighted local coefficients at low resolution, upsamples them, and generates a high-resolution output.
- AG-Net uses M-Net as its backbone and replaces conventional skip connections and upsampling with an attention guided filter.
- Attention Guided Filter: The filter takes guidance feature map I and filtering feature map O, then produces high-resolution output ˜O using attention feature map T.
- Attention Guided Filter: It downsamples I to Il, estimates coefficients Al and Bl by minimizing reconstruction error against O, then upsamples them to Ah and Bh.
- Attention Guided Filter: The local coefficients are computed within windows of radius r using attention weights Ti and regularization parameter λ.
- Attention Guided Filter: Overlapping-window outputs are averaged, and the upsampled coefficients generate ˜O through element-wise multiplication with I plus Bh.
- Attention Block: The attention block transforms O and Il, combines them by element-wise addition with ReLU, and produces an attention map through a further convolution.
3 Experiments
Experiments evaluate AG-Net on retinal vessel segmentation and optic disc/cup segmentation across DRIVE, CHASE DB1, and ORIGA. Results indicate that guided filtering preserves structural information, while attention improves discrimination and segmentation boundaries.
- Experimental Setup: AG-Net is evaluated on vessel segmentation and optic disc/cup segmentation using DRIVE, CHASE DB1, and ORIGA datasets.The vessel experiments include DRIVE and CHASE DB1, while optic disc/cup segmentation uses ORIGA.
- Vessel Segmentation: GF-Net outperforms original M-Net on DRIVE, supporting guided filtering over skip connections for transferring structural information.The comparison is explicitly presented as evidence for the guided filter's structural-information transfer.
- Vessel Segmentation: 0.0010, 0.0019, 0.0205 and 0.0126 are AG-Net's gains over GF-Net in Acc, AUC, Sen and IOU, respectively, on DRIVE.The reported improvements correspond in order to Accuracy, Area Under ROC, Sensitivity, and intersection-over-union.
- Vessel Segmentation: AG-Net produces more precise vessel boundaries than M-Net+GF, while GF-Net produces clearer boundaries than M-Net.The visual comparison attributes clearer boundaries to guided filtering and foreground/background discrimination to attention.
- Optic Disc/Cup Segmentation: On ORIGA, the method outperforms all compared state-of-the-art optic disc and/or cup segmentation algorithms on the reported overlapping-error criteria.AG-MNet also performs much better than original M-Net under the same setting.
4 Conclusions
The paper proposes an attention guided filter as a structure-sensitive expanding path within M-Net. Across two retinal segmentation tasks, the method is reported to improve information fusion while highlighting foreground and reducing background effects.
- 4 Conclusions: AG-Net replaces skip connections and upsampling with an attention guided filter that transfers structural information and improves information fusion.The attention mechanism highlights foreground regions and reduces the effect of background.
- 4 Conclusions: Experiments on vessel segmentation and optic disc/cup segmentation demonstrate the effectiveness of the proposed method.