Fig 1.
Overview of MFI-Net segmentation model for retinal vessel.
Fig 2.
The structure of residual unit.
Fig 3.
Sample images of DRIVE, CHASE DB1 and STARE datasets.
Fig 4.
(a) is the original medical image. (b–d) are visualizations of red, green, and blue channels, respectively. Results of each preprocessing strategy. (e) is the image after grayscale processing. (f) is the image after data normalization. (g) is the image after CLAHE operation. (h) is the image after Gaussian correction.
Table 1.
Ablation experiment results.
Fig 5.
PR and ROC curves of each ablation structure.
Table 2.
P-value analysis results among various ablation models on different dataset.
Fig 6.
Segmentation result of ablation experiment.
The areas that require special attention are zoomed in, and areas with obvious differences are marked with red rectangles.
Table 3.
Segmentation results of different models on the DRIVE dataset.
Table 4.
Segmentation results of different models on the CHASE_DB1 dataset.
Table 5.
Segmentation results of different models on the STARE dataset.
Fig 7.
Comparison of segmentation results of MFI-Net(ours), UNet++ and AA-UNet.
The red rectangles mark areas with obvious differences.
Table 6.
Details of the parameter amount and time cost of different models.
Fig 8.
F1 performance curve on three datasets.
The abscissa represents the sample picture number in the dataset, the ordinate represents the F1 for semantic segmentation of the image.
Fig 9.
Segmentation result of cross-validate experiment.
The first column is the original image; The third to fifth columns are the results of segmenting the three data sets using the model files trained on DRIVE, CHASE DB1 and STARE by MFI-Net.
Table 7.
Cross test results.
Table 8.
Results of cross-testing on DRIVE and STARE dataset.