Abstract
Image dehazing deals with the removal of undesired loss of visibility in outdoor images due to the presence of fog. Retinex is a color vision model mimicking the ability of the Human Visual System to robustly discount varying illuminations when observing a scene under different spectral lighting conditions. Retinex has been widely explored in the computer vision literature for image enhancement and other related tasks. While these two problems are apparently unrelated, the goal of this work is to show that they can be connected by a simple linear relationship. Specifically, most Retinex-based algorithms have the characteristic feature of always increasing image brightness, which turns them into ideal candidates for effective image dehazing by directly applying Retinex to a hazy image whose intensities have been inverted. In this paper, we give theoretical proof that Retinex on inverted intensities is a solution to the image dehazing problem. Comprehensive qualitative and quantitative results indicate that several classical and modern implementations of Retinex can be transformed into competing image dehazing algorithms performing on pair with more complex fog removal methods, and can overcome some of the main challenges associated with this problem.

Abstract
Deep learning models have been successfully used in medical image analysis problems but they require a large amount of labeled images to obtain good performance. However, such large labeled datasets are costly to acquire. Active learning techniques can be used to minimize the number of required training labels while maximizing the model's performance. In this work, we propose a novel sampling method that queries the unlabeled examples that maximize the average distance to all training set examples in a learned feature space. We then extend our sampling method to define a better initial training set, without the need for a trained model, by using Oriented FAST and Rotated BRIEF (ORB) feature descriptors. We validate MedAL on 3 medical image datasets and show that our method is robust to different dataset properties. MedAL is also efficient, achieving 80% accuracy on the task of Diabetic Retinopathy detection using only 425 labeled images, corresponding to a 32% reduction in the number of required labeled examples compared to the standard uncertainty sampling technique, and a 40% reduction compared to random sampling.

Abstract
Diabetic Retinopathy (DR) is one of the leading causes of preventable blindness in the developed world. With the increasing number of diabetic patients there is a growing need of an automated system for DR detection. We propose EyeWeS, a method that not only detects DR in eye fundus images but also pinpoints the regions of the image that contain lesions, while being trained with image labels only. We show that it is possible to convert any pre-trained convolutional neural network into a weakly-supervised model while increasing their performance and efficiency. EyeWeS improved the results of Inception V3 from 94:9% Area Under the Receiver Operating Curve (AUC) to 95:8% AUC while maintaining only approximately 5% of the Inception V3's number of parameters. The same model is able to achieve 97:1% AUC in a cross-dataset experiment.

Abstract
The automatic differentiation of retinal vessels into arteries and veins (A/V) is a highly relevant task within the field of retinal image analysis. however, due to limitations of retinal image acquisition devices, specialists can find it impossible to label certain vessels in eye fundus images. In this paper, we introduce a method that takes into account such uncertainty by design. For this, we formulate the A/V classification task as a four-class segmentation problem, and a Convolutional Neural Network is trained to classify pixels into background, A/V, or uncertain classes. The resulting technique can directly provide pixelwise uncertainty estimates. In addition, instead of depending on a previously available vessel segmentation, the method automatically segments the vessel tree. Experimental results show a performance comparable or superior to several recent A/V classification approaches. In addition, the proposed technique also attains state-of-the-art performance when evaluated for the task of vessel segmentation, generalizing to data that, was not used during training, even with considerable differences in terms of appearance and resolution.

Abstract
Lung volume segmentation is a key step in the design of Computer-Aided Diagnosis systems for automated lung pathology analysis. However, isolating the lung from CT volumes can he a challenging process due to considerable deformations and the potential presence of pathologies. Convolutional Neural Networks (CNN) are effective tools for modeling the spatial relationship between lung voxels. Unfortunately, they typically require large quantities of annotated data, and manually delineating the lung from volumetric CT scans can he a cumbersome process. We propose to train a 3D CNN to solve this task based on semi-automatically generated annotations. For this, we introduce an extension of the well-known V-Net architecture that can handle higher dimensional input data. Even if the training set labels are noisy and contain errors, our experiments show that it is possible to learn to accurately segment the lung relying on them. Numerical comparisons on an external test set containing lung segmentations provided by a medical expert demonstrate that the proposed model generalizes well to new data, reaching an average 98.7% Dice coefficient. The proposed approach results in a superior performance with respect to the standard V-Net model, particularly on the lung boundary.

Abstract
Visual exploration of the larynx represents a relevant technique for the early diagnosis of laryngeal disorders. However, visualizing an endoscopy for finding abnormalities is a time-consuming process, and for this reason much research has been dedicated to the automatic analysis of endoscopic video data. In this work we address the particular task of discriminating among informative laryngoscopic frames and those that carry insufficient diagnostic information. In the latter case, the goal is also to determine the reason for this lack of information. To this end, we analyze the possibility of training three different state-of-the-art Convolutional Neural Networks, but initializing their weights from configurations that have been previously optimized for solving natural image classification problems. Our findings show that the simplest of these three architectures not only is the most accurate (outperforming previously proposed techniques), but also the fastest and most efficient, with the lowest inference time and minimal memory requirements, enabling real-time application and deployment in portable devices.