Abstract
Traditionally, convolutional neural networks are trained for semantic segmentation by having an image given as input and the segmented mask as output. In this work, we propose a neural network trained by being given an image and mask pair, with the output being the quality of that pairing. The segmentation is then created afterwards through backpropagation on the mask. This allows enriching training with semi-supervised synthetic variations on the ground-truth. The proposed iterative segmentation technique allows improving an existing segmentation or creating one from scratch. We compare the performance of the proposed methodology with state-of-the-art deep architectures for image segmentation and achieve competitive results, being able to improve their segmentations.

Abstract
Facial expression recognition (FER) is currently one of the most active research topics due to its wide range of applications in the human-computer interaction field. An important part of the recent success of automatic FER was achieved thanks to the emergence of deep learning approaches. However, training deep networks for FER is still a very challenging task, since most of the available FER data sets are relatively small. Although transfer learning can partially alleviate the issue, the performance of deep models is still below of its full potential as deep features may contain redundant information from the pre-trained domain. Instead, we propose a novel end-to-end neural network architecture along with a well-designed loss function based on the strong prior knowledge that facial expressions are the result of the motions of some facial muscles and components. The loss function is defined to regularize the entire learning process so that the proposed neural network is able to explicitly learn expression-specific features. Experimental results demonstrate the effectiveness of the proposed model in both lab-controlled and wild environments. In particular, the proposed neural network provides quite promising results, outperforming in most cases the current state-of-the-art methods.

Abstract
Interpretability is a fundamental property for the acceptance of machine learning models in highly regulated areas. Recently, deep neural networks gained the attention of the scientific community due to their high accuracy in vast classification problems. However, they are still seen as black-box models where it is hard to understand the reasons for the labels that they generate. This paper proposes a deep model with monotonic constraints that generates complementary explanations for its decisions both in terms of style and depth. Furthermore, an objective framework for the evaluation of the explanations is presented. Our method is tested on two biomedical datasets and demonstrates an improvement in relation to traditional models in terms of quality of the explanations generated.

Abstract
Two limitations hamper performance of deep architectures for classification and/or detection in medical imaging: (i) the small amount of available data, and (ii) the class imbalance scenario. While millions of labeled images are available today to build classification tools for natural scenes, the amount of available annotated data for automatic breast cancer screening is limited to a few thousand images, at best. We address these limitations with a method for data augmentation, based on the introduction of random elastic deformations on images of mammograms. We validate this method on three publicly available datasets. Our proposed Convolutional Neural Network (CNN) archi-tecture is trained for mass classification - in a conventional way - , and then used in the more interesting problem of mass detection in full mammograms by transforming the CNN into a Fully Convolutional Network (FCN).

Abstract
Facial recognition under uncontrolled acquisition environments faces major challenges that limit the deployment of real-life systems. The use of 2.5D information can be used to improve discriminative power of such systems in conditions where RGB information alone would fail. In this paper we propose a multimodal extension of a previous work, based on SIFT descriptors of RGB images, integrated with LBP information obtained from depth scans, modeled by an hierarchical framework motivated by principles of human cognition. The framework was tested on EURECOM dataset and proved that the inclusion of depth information improved significantly the results in all the tested conditions, compared to independent unimodal approaches.

Abstract
Infantile Hemangiomas (IH) make up the most common type of benign vascular tumors affecting children. They can grow for several months until beginning to involute. In present-day clinical practice there's no objective monitoring protocol. For more objective measures, an automatic evaluation system (CAD system) is needed to aid clinicians in assessing the effectiveness of a given patient's response to a treatment. One of the stages of these systems is the lesion segmentation. This work addresses the automatic segmentation of lesions in IH. Acknowledging that the methods in the literature for IH lesion segmentation lag behind the state-of-the-art in the image segmentation community, we conduct a comparison of various methodologies for the segmentation of the IH, including both shallow and deep methodologies. Acknowledging the lack of data in the field for a robust learning of deep models, we also evaluate transfer learning techniques to benefit from knowledge extracted in other skin lesions. The best results were obtained with the shortest path method and a multiscale convolutional neural network that merges two pipelines working at different scales. Although promising, the results put in evidence the need for better databases, collected under suitable acquisition protocols.