Fig 1.
The overview of the proposed HarDenseNet.
HarDNet Block was used to replace the original convolution operation in the down-sampling process, RFB(Receptive Field Block Module) block was used for skip connections, and Dense Aggregation was used in the up-sampling process.
Fig 2.
HarDNet block overview.
Fig 3.
Receptive field block overview.
Fig 4.
Aggregation module overview.
Fig 5.
A demonstration of some examples from the DRIVE and CHASE DB1 datasets experiment.
The three columns on the left are from the DRIVE dataset and the three columns on the right are from the CHASE DB1 datasets. The first line are the color fundus images from the two data sets, the second line is the corresponding ground truth, and the third line is the segmentation results based on the proposed method.
Fig 6.
Comparison experiment between the proposed method and U-Net on DRIVE dataset, and the corresponding area is enlarged: The first line are the color fundus image from the Drive dataset; The second line are the corresponding manual annotations; The third line are the proposed segmentation results; And the fourth line are the U-Net segmentation results.
Fig 7.
Comparison of the proposed HarDNet and U-Net on CHASE DB1 dataset, and the corresponding area is enlarged: First line are color fundus images; The second line are ground truth; The third line are segmentation result by The third line is; The last line are segmentation result by U-Net.
Table 1.
Results of different methods on DRIVE and CHASE DB1.
Table 2.
P-values for performance differences between the proposed method and baselines.
Fig 8.
The validation accuracy curves obtained during the training processes of the two datasets, DRIVE and CHASE DB1.
Fig 9.
The ROC curve obtained from the DRIVE dataset.
Fig 10.
The ROC curve obtained from the CHASE DB1 dataset.
Table 3.
Ablation experiments on the DRIVE dataset.
Table 4.
Comprehensive parameters comparative analysis of the proposed network with Unet.
Fig 11.
A comparison between the pre – processed images of two datasets and their corresponding gold standards is presented.
The left – most three columns of images are derived from the DRIVE dataset, while the right – most three columns are from the CHASE DB1 dataset. In sequence from top to bottom are the original images, the pre – processed images, and the gold standards.