Search

[EN] Chest X-ray images are useful for early COVID-19 diagnosis with the advantage that X-ray devices are already available in health centers and images are obtained immediately. Some datasets containing X-ray images with cases (pneumonia or COVID-19) and controls have been made available to develop machine-learning-based methods to aid in diagnosing the disease. However, these datasets are mainly composed of different sources coming from pre-COVID-19 datasets and COVID-19 datasets. Particularly, we have detected a significant bias in some of the released datasets used to train and test diagnostic systems, which might imply that the results published are optimistic and may overestimate the actual predictive capacity of the techniques proposed. In this article, we analyze the existing bias in some commonly used datasets and propose a series of preliminary steps to carry out before the classic machine learning pipeline in order to detect possible biases, to avoid them if possible and to report results that are more representative of the actual predictive power of the methods under analysis.

[EN] Background and Objective: Breast cancer is the most frequent cancer in women. The Spanish healthcare network established population-based screening programs in all Autonomous Communities, where mammograms of asymptomatic women are taken with early diagnosis purposes. Breast density assessed from digital mammograms is a biomarker known to be related to a higher risk to develop breast cancer. It is thus crucial to provide a reliable method to measure breast density from mammograms. Furthermore the complete automation of this segmentation process is becoming fundamental as the amount of mammograms increases every day. Important challenges are related with the differences in images from different devices and the lack of an objective gold standard. This paper presents a fully automated framework based on deep learning to estimate the breast density. The framework covers breast detection, pectoral muscle exclusion, and fibroglandular tissue segmentation. Methods: A multi-center study, composed of 1785 women whose "for presentation" mammograms were segmented by two experienced radiologists. A total of 4992 of the 6680 mammograms were used as training corpus and the remaining (1688) formed the test corpus. This paper presents a histogram normalization step that smoothed the difference between acquisition, a regression architecture that learned segmentation parameters as intrinsic image features and a loss function based on the DICE score. Results: The results obtained indicate that the level of concordance (DICE score) reached by the two radiologists (0.77) was also achieved by the automated framework when it was compared to the closest breast segmentation from the radiologists. For the acquired with the highest quality device, the DICE score per acquisition device reached 0.84, while the concordance between radiologists was 0.76. Conclusions: An automatic breast density estimator based on deep learning exhibits similar performance when compared with two experienced ra

[EN] Inspecting a 3D object which shape has elastic manufacturing tolerances in order to find defects is a challenging and time-consuming task. This task usually involves humans, either in the specification stage followed by some automatic measurements, or in other points along the process. Even when a detailed inspection is performed, the measurements are limited to a few dimensions instead of a complete examination of the object. In this work, a probabilistic method to evaluate 3D surfaces is presented. This algorithm relies on a training stage to learn the shape of the object building a statistical shape model. Making use of this model, any inspected object can be evaluated obtaining a probability that the whole object or any of its dimensions are compatible with the model, thus allowing to easily find defective objects. Results in simulated and real environments are presented and compared to two different alternatives.

[EN] Pedestrian detection is a particular case of object detection that helps to reduce accidents in advanced driver-assistance systems and autonomous vehicles. It is not an easy task because of the variability of the objects and the time constraints. A performance comparison of object detection methods, including both GPU and non-GPU implementations over a variety of on-road specific databases, is provided. Computer vision multi-class object detection can be integrated on sensor fusion modules where recall is preferred over precision. For this reason, ad hoc training with a single class for pedestrians has been performed and we achieved a significant increase in recall. Experiments have been carried out on several architectures and a special effort has been devoted to achieve a feasible computational time for a real-time system. Finally, an analysis of the input image size allows to fine-tune the model and get better results with practical costs.

[EN] A precise calibration in multi-view camera environments allows to perform accurate 3D object reconstruction, precise tracking of objects and accurate pose estimation. Those techniques are of high value in the industry today in fields as quality control or automation. In the present work, an improvement of a simple existing multi-view camera calibration method is presented. The improved method employs a specially developed reference token to overcome some issues in the original algorithm. We prove that the new method overcomes those problems thus attaining a higher accuracy while keeping the process simple and the implementation costs low. This last aspects makes the method interesting for the industry but specially suitable for SMEs typical in traditional sectors.

[EN] Vulnerable Road User (VRU) detection is a major application of object detection with the aim of helping reduce accidents in advanced driver-assistance systems and enabling the development of autonomous vehicles. Due to intrinsic complexity present in computer vision and to limitations in processing capacity and bandwidth, this task has not been completely solved nowadays. For these reasons, the well established YOLOv3 net and the new YOLOv4 one are assessed by training them on a huge, recent on-road image dataset (BDD100K), both for VRU and full on-road classes, with a great improvement in terms of detection quality when compared to their MS-COCO-trained generic correspondent models from the authors but with negligible costs in forward pass time. Additionally, some models were retrained when replacing the original Leaky ReLU convolutional activation functions from original YOLO implementation with two cutting-edge activation functions: the self-regularized non-monotonic function (MISH) and its self-gated counterpart (SWISH), with significant improvements with respect to the original activation function detection performance. Additionally, some trials were carried out including recent data augmentation techniques (mosaic and cutmix) and some grid size configurations, with cumulative improvements over the previous results, comprising different performance-throughput trade-offs.

[EN] Background The breast dense tissue percentage on digital mammograms is one of the most commonly used markers for breast cancer risk estimation. Geometric features of dense tissue over the breast and the presence of texture structures contained in sliding windows that scan the mammograms may improve the predictive ability when combined with the breast dense tissue percentage. Methods A case/control study nested within a screening program covering 1563 women with craniocaudal and mediolateral-oblique mammograms (755 controls and the contralateral breast mammograms at the closest screening visit before cancer diagnostic for 808 cases) aging 45 to 70 from Comunitat Valenciana (Spain) was used to extract geometric and texture features. The dense tissue segmentation was performed using DMScan and validated by two experienced radiologists. A model based on Random Forests was trained several times varying the set of variables. A training dataset of 1172 patients was evaluated with a 10-stratified-fold cross-validation scheme. The area under the Receiver Operating Characteristic curve (AUC) was the metric for the predictive ability. The results were assessed by only considering the output after applying the model to the test set, which was composed of the remaining 391 patients. Results The AUC score obtained by the dense tissue percentage (0.55) was compared to a machine learning-based classifier results. The classifier, apart from the percentage of dense tissue of both views, firstly included global geometric features such as the distance of dense tissue to the pectoral muscle, dense tissue eccentricity or the dense tissue perimeter, obtaining an accuracy of 0.56. By the inclusion of a global feature based on local histograms of oriented gradients, the accuracy of the classifier was significantly improved (0.61). The number of well-classified patients was improved up to 236 when it was 208. Conclusion Relative geometric features of dense tissue over the breast and his

[EN] This work presents a 3D scanner able to reconstruct a complete object without occlusions, including its surface appearance. The technique presents a number of differences in relation to current scanners: it does not require mechanical handling like robot arms or spinning plates, it is free of occlusions since the scanned part is not resting on any surface and, unlike stereo-based methods, the object does not need to have visual singularities on its surface. This system, among other applications, allows its integration in production lines that require the inspection of a large volume of parts or products, especially if there is an important variability of the objects to be inspected, since there is no mechanical manipulation. The scanner consists of a variable number of industrial quality cameras conveniently distributed so that they can capture all the surfaces of the object without any blind spot. The object is dropped through the common visual field of all the cameras, so no surface or tool occludes the views that are captured simultaneously when the part is in the center of the visible volume. A carving procedure that uses the silhouettes segmented from each image gives rise to a volumetric representation and, by means of isosurface generation techniques, to a 3D model. These techniques have certain limitations on the reconstruction of object regions with particular geometric configurations. Estimating the inherent maximum error in each area is important to bound the precision of the reconstruction. A number of experiments are presented reporting the differences between ideal and reconstructed objects in the system.

We use Weighted Finite-State Transducers (WFSTs) to represent the different sources of information available: the initial hypotheses, the possible errors, the constraints imposed by the task (interaction language) and the user input. The fusion of these models to find the most probable output string can be performed efficiently by using carefully selected transducer operations. The proposed system initially suggests an output based on the set of hypotheses, possible errors and Constraint Models. Then, if human intervention is needed, a multimodal approach, where the user input is combined with the aforementioned models, is applied to produce, with a minimum user effort, the desired output. This approach offers the practical advantages of a de-coupled model (e.g. input-system + parameterized rules + post-processor), keeping at the same time the error-recovery power of an integrated approach, where all the steps of the process are performed in the same formal machine (as in a typical HMM in speech recognition) to avoid that an error at a given step remains unrecoverable in the subsequent steps. After a presentation of the theoretical basis of the proposed multi-source information system, its application to two real world problems, as an example of the possibilities of this architecture, is addressed. The experimental results obtained demonstrate that a significant user effort can be saved when using the proposed procedure. A simple demonstration, to better understand and evaluate the proposed system, is available on the web https://demos.iti.upv.es/hi/. (C) 2015 Elsevier B.V. All rights reserved.

An OCR process is often followed by the application of a language model to find the best transformation of an OCR hypothesis into a string compatible with the constraints of the document, field or item under consideration. The cost of this transformation can be taken as a confidence value and compared to a threshold to decide if a string is accepted as correct or rejected in order to satisfy the need for bounding the error rate of the system. Widespread tools like ROC, precision-recall, or error-reject curves, are commonly used along with fixed thresholding in order to achieve that goal. However, those methodologies fail when a test sample has a confidence distribution that differs from the one of the sample used to train the system, which is a very frequent case in post-processed OCR strings (e.g., string batches showing particularly careful handwriting styles in contrast to free styles). In this paper, we propose an adaptive method for the automatic estimation of the rejection threshold that overcomes this drawback, allowing the operator to define an expected error rate within the set of accepted (non-rejected) strings of a complete batch of documents (as opposed to trying to establish or control the probability of error of a single string), regardless of its confidence distribution. The operator (expert) is assumed to know the error rate that can be acceptable to the user of the resulting data. The proposed system transforms that knowledge into a suitable rejection threshold. The approach is based on the estimation of an expected error vs. transformation cost distribution. First, a model predicting the probability of a cost to arise from an erroneously transcribed string is computed from a sample of supervised OCR hypotheses. Then, given a test sample, a cumulative error vs. cost curve is computed and used to automatically set the appropriate threshold that meets the user-defined error rate on the overall sample. The results of experiments on batches coming from

The task of breast density quantification is becoming increasingly relevant due to its association with breast cancer risk. In this work, a semi-automated and a fully automated tools to assess breast density from full-field digitized mammograms are presented. The first tool is based on a supervised interactive thresholding procedure for segmenting dense from fatty tissue and is used with a twofold goal: for assessing mammographic density(MD) in a more objective and accurate way than via visual-based methods and for labeling the mammograms that are later employed to train the fully automated tool. Although most automated methods rely on supervised approaches based on a global labeling of the mammogram, the proposed method relies on pixel-level labeling, allowing better tissue classification and density measurement on a continuous scale. The fully automated method presented combines a classification scheme based on local features and thresholding operations that improve the performance of the classifier. A dataset of 655 mammograms was used to test the concordance of both approaches in measuring MD. Three expert radiologists measured MD in each of the mammograms using the semi-automated tool (DM-Scan). It was then measured by the fully automated system and the correlation between both methods was computed. The relation between MD and breast cancer was then analyzed using a case-control dataset consisting of 230 mammograms. The Intraclass Correlation Coefficient (ICC) was used to compute reliability among raters and between techniques. The results obtained showed an average ICC = 0.922 among raters when using the semi-automated tool, whilst the average correlation between the semi-automated and automated measures was ICC = 0.838. In the case-control study, the results obtained showed Odds Ratios (OR) of 1.38 and 1.50 per 10% increase in MD when using the semi-automated and fully automated approaches respectively. It can therefore be concluded that the automated and sem

Boletín Informativo del Instituto Tecnológico de Informática, dedicado a las Tecnologías de la información y las Comunicaciones