With no other methodological innovation than a carefully designed training, our ResNet model achieved an AUC = 0.955 (0.953 - 0.956) on a combined test set of 61007 test images from different public datasets, which is in line or even better than what other more complex deep learning models reported in the literature.
We train the first deep learning model for segmenting brain arteries from 3D rotational angiographies in cases with brain arterio-venous malformations.
We experimentally validate whether using coarse-to-fine models instead of one-stage models is appropriate or not for segmenting the optic disc and the optic cup in color fundus images. We observed that models trained with the right amount of data can perform much better than coarse-to-find approaches.
We posed a multiclass segmentation task as a single multitask model with binary segmentation targets. Our results indicate that this approach might be useful to deal with "sandwiched" structures.
We introduce an augmented target loss function framework for photoreceptor layer segmentation that penalizes errors in the central area of each B-scan. It allows to significantly improve performance with respect to the standard loss functions.
We introduced a multitask learning y-shaped neural network that simultaneously segment the FAZ in FA images and predict a distance map. This extra branch aids to improve results in clinical routine images.
We developed a Bayesian U-Net model for photoreceptor layer segmentation in OCT that predicts epistemic uncertainty maps highlighting potential areas of error in the segmentation.
We used CycleGANs to translate OCT images from one vendor to another. This approach allows us to increase the performance of fluid segmentation models trained on one vendor and evaluated on another.
We designed a method to summarize hemodynamic parameters obtained by 0D simulations so that they can be applied for glaucoma detection. We observed certain correlation between glaucoma and these hemodynamic features.