Fig 1.
The proposed architecture of PCAT-UNet.
Our PCAT-UNet is composed of encoder and decoder constructed by PCAT blocks, convolutional branch constructed FGAM, skip connection and right output layer. In our network, the PCAT block is used as a structurally sensitive skip connection to achieve better information fusion. Finally, the side output layer uses the fused enhanced feature map to predict the vessels segmentation map of each layer along the decoder path.
Fig 2.
Structure diagram of PCAT block and PCA.
(a) two consecutive PCAT blocks.CPCA and IPCA are multi-head self-attention modules with cross and inner patching configurations, respectively. (b)The detailed structure of the PCA, a MHSA based on convolution.
Fig 3.
Structure comparison of EPSA module and FGAM.
(a) EPSA module; (b) FGAM.
Table 1.
The specific information of DRIVE, STARE and CHASE_DB1 datasets.
Fig 4.
Retinal vessels images of the dataset (line 1), corresponding retinal vessels images generated by histogram stretching enhancement (line 2) and corresponding FOV mask (line 3); (a) DRIVE (b) STARE (c) CHASE_DB1.
Table 2.
Performance comparison of different segmentation methods on the DRIVE dataset.
Table 3.
Performance comparison of different segmentation methods on the STARE dataset.
Table 4.
Performance comparison of different segmentation methods on the CHASE_DB1 dataset.
Fig 5.
ROC Curves on (a) DRIVE (b) STARE (c) CHASE_DB1.
Fig 6.
Comparison of example vessel segmentation results on DRIVE dataset: (a) original retinal images; (b) ground truths; (c) segmentation results for UNet; (d) segmentation results for PCAT-UNet. The first row of the image is the whole image, and the second row is the zoomed in area of the marked red border and blue border in the image.
Fig 7.
Comparison of example vessel segmentation results on STARE dataset: (a) original retinal images; (b) ground truths; (c) segmentation results for UNet; (d) segmentation results for PCAT-UNet.
The first row of the image is the whole image, and the second row is the zoomed in area of the marked red border and blue border in the image.
Fig 8.
Comparison of example vessel segmentation results on CHASE_DB1 dataset: (a) original retinal images; (b) ground truths; (c) segmentation results for UNet; (d) segmentation results for PCAT-UNet.
The first row of the image is the whole image, and the second row is the zoomed in area of the marked red border and blue border in the image.
Table 5.
Ablation study results on DRIVE dataset.
Table 6.
Ablation study results on STARE dataset.
Table 7.
Ablation study results on CHASE_DB1 dataset.
Table 8.
Quantitative comparison of parameter and time consumption.
Table 9.
Performance comparison of EPSA module and FGAM on three datasets.