Abstract
The segmentation of blood vessels in medical images has been heavily studied, given its impact in several clinical practices. Deep Learning methods have been applied to supervised segmentation of blood vessels, mainly the retinal ones due to the availability of manual annotations. Despite their success, they typically minimize the Binary Cross Entropy loss, which does not penalize topological mistakes. These errors are relevant in graph-like structures such as blood vessel trees, as a missing segment or an inadequate merging or splitting of branches, may severely change the topology of the network and put at risk the extraction of vessel pathways and their characterization. In this paper, we propose an end-to-end network design comprising a cascade of a typical segmentation network and a Variational Auto-Encoder which, by learning a rich but compact latent space, is able to correct many topological incoherences. Our experiments in three of the most commonly used retinal databases, DRIVE, STARE, and CHASEDB1, show that the proposed model effectively learns representations inducing better segmentations in terms of topology, without hurting the usual pixel-wise metrics.

Abstract
As a key technology to help bridging the gap between deaf and hearing people, sign language recognition (SLR) has become one of the most active research topics in the human-computer interaction field. Although several SLR methodologies have been proposed, the development of a real-world SLR system is still a very challenging task. One of the main challenges is related to the large intersigner variability that exists in the manual signing process of sign languages. To address this problem, we propose a novel end-to-end deep neural network that explicitly models highly discriminative signer-independent latent representations from the input data. The key idea of our model is to learn a distribution over latent representations, conditionally independent of signer identity. Accordingly, the learned latent representations will preserve as much information as possible about the signs, and discard signer-specific traits that are irrelevant for recognition. By imposing such regularization in the representation space, the result is a truly signer-independent model which is robust to different and new test signers. The experimental results demonstrate the effectiveness of the proposed model in several SLR databases.

Abstract
Convolutional Neural Networks (CNN) have become the gold standard in many visual recognition tasks including medical applications. Due to their high variance, however, these models are prone to over-fit the data they are trained on. To mitigate this problem, one of the most common strategies, is to perform data augmentation. Rotation, scaling and translation are common operations. In this work we propose an alternative method to rotation-based data augmentation where the rotation transformation is performed inside the CNN architecture. In each training batch the weights of all convolutional layers are rotated by the same random angle. We validate our proposed method empirically showing its usefulness under different scenarios.

Abstract
Falls are very rare and extremely difficult to acquire in free living conditions. Due to this, most of prior work on fall detection has focused on simulated datasets acquired in scenarios that mimic the real-world context, however, the validation of systems trained with simulated falls remains unclear. This work presents a transfer learning approach for combining a dataset of simulated falls and non-falls, obtained from young volunteers, with the real-world FARSEEING dataset, in order to train a set of supervised classifiers for discriminating between falls and non-falls events. The objective is to analyze if a combination of simulated and real falls could enrich the model. In the real-world, falls are a sporadic event, which results in imbalanced datasets. In this work, several methods for imbalance learning were employed: SMOTE, Balance Cascade and Ranking models. The Balance Cascade obtained less misclassifications in the validation set.There was an improvement when mixing the real falls and simulated non-falls compared to the case when only simulated falls were used for training. When testing with a mixed set with real falls and simulated non-falls, it is even more important to train with a mixed set. Moreover, it was possible to onclude that a model trained with simulated falls generalize better when tested with real falls, than the opposite. The overall accuracy obtained for the combination of different datasets were above 95 %. © 2018 IEEE.

Abstract
Ordinal arrangement of objects is a common property in biomedical images. Traditional methods to deal with semantic image segmentation in this setting are ad-hoc and application specific. In this paper, we propose ordinal-aware deep learning architectures for image segmentation that enforce pixelwise consistency by construction. We validated the proposed architectures on several real-life biomedical datasets and achieved competitive results in all cases. © 2018 IEEE.

Abstract
We propose a framework to model the distribution of sequential data coming from a set of entities connected in a graph with a known topology. The method is based on a mixture of shared hidden Markov models (HMMs), which are jointly trained in order to exploit the knowledge of the graph structure and in such a way that the obtained mixtures tend to be sparse. Experiments in different application domains demonstrate the effectiveness and versatility of the method.