upd paper

paper/paper.md

@@ -45,7 +45,7 @@ The Fundus Image Toolbox was developed to address this need within the medical i
 # Tools
 The main functionalities of the Fundus Image Toolbox are:
 
-- Quality prediction (\autoref{fig:example}a). We trained an ensemble of ResNets and EfficientNets on the combined DeepDRiD and DrimDB datasets [@deepdrid;@drimdb] to predict the gradeability of fundus images. The datasets are publicly available and comprise images of retinas with diabetic retinopathy, healthy retinas and outliers such as outer eye images. The model ensemble achieved an accuracy of 0.78 and an area under the receiver operating characteristic curve (AUROC) of 0.84 on a DeepDRiD test split, surpassing the previous best model evaluated on DeepDRiD [@tummala2023]. Further, on a DrimDB test split, our model acchieved an accuracy and AUROC of 1.0 and 1.0, respectively.
+- Quality prediction (\autoref{fig:example}a). We trained an ensemble of ResNets and EfficientNets on the combined DeepDRiD and DrimDB datasets [@deepdrid;@drimdb] to predict the gradeability of fundus images. The datasets are publicly available and comprise images of retinas with diabetic retinopathy, healthy retinas and outliers such as outer eye images. The model ensemble achieved an accuracy of 0.78 and an area under the receiver operating characteristic curve (AUROC) of 0.84 on a DeepDRiD test split, surpassing the previous best model evaluated on DeepDRiD [@tummala2023]. Further, on a DrimDB test split, the accuracy and AUROC of our model were 1.0 and 1.0, respectively.
 - Fovea and optic disc localization (\autoref{fig:example}b). The center coordinates of the fovea and optic disc can be predicted using a multitask EfficientNet model. We trained the model on the combined ADAM, REFUGE and IDRID datasets which include images from eyes with age-related macular degeneration, glaucoma, diabetic retinopathy and healthy retinas [@adam;@refuge;@idrid]. All datasets are publicly available. On our test split, the model achieved an average distance to the fovea and optic disc targets of 0.88 % of the image size. This corresponds to a mean distance of 3.08 pixels in the 350 x 350 pixel images used for training and testing.
 - Vessel segmentation (\autoref{fig:example}c). The segmentation method produces a mask of blood vessels in a fundus image using an ensemble of FR-UNets. The ensemble achieved an average Dice score of 0.887 on the test split of the FIVES dataset [@koehler2024]. FIVES includes images with age-related macular degeneration, glaucoma, diabetic retinopathy and healthy retinas [@fives].
 - Registration (\autoref{fig:example}d). Two fundus images of the same eye can be aligned using SuperRetina [@liu2022]. The deep learning based model detects key points on the vessel trees of the two images and matches them. This results in a registered version of the second image that is aligned with the first. SuperRetina produced registrations of at least acceptable quality in 98.5 % of the cases on the test split of the FIRE dataset [@fire].