Figure 4.

Typical examples of AOSLO-net outperforming nnUNet on detecting (a) the MA body and (b) parenting vessels of MAs when they are trained using normal mask set (n-Mask). First column: enhanced AOSLO images. Second column: normal masks used to train the models. Third column: segmentation results of nnUNet. Fourth column: segmentation results of AOSLO-net. Numbers in images are the Dice scores. (a) For images with IDs 036, 043, 048, 062, and 092, the AOSLO-net is able to detect MAs, which are missed by nnUNet (marked in red circles). For images with IDs 013 and 048, the AOSLO-net can better reconstruct the full shape of MAs compared to those of nnUNet (marked in yellow circles). We also note that in cases of IDs 013, 043 and 092, the segmentation models can even detect some potential MAs, which are not marked in the masks. (b) For images with IDs 015, 055, 082, 088, 094, and 102, the AOSLO-net is capable of detecting both feeding and draining vessels connected to the MAs, whereas the nnUNet may miss one or both of the parenting vessels (marked in yellow circles). These results show that the AOSLO-net is more reliable in detecting MA parenting vessels from the input images, which are essential in the MA classification—a downstream task for disease diagnosis.