Deep Siamese Metric Learning: A Highly Scalable Approach to Searching Unordered Sets of Trajectories

This work proposes metric learning for fast similarity-based scene retrieval of unstructured ensembles of trajectory data from large databases. We present a novel representation learning approach using Siamese Metric Learning that approximates a distance preserving low-dimensional representation and that learns to estimate reasonable solutions to the assignment problem. To this end, we employ a Temporal Convolutional Network architecture that we extend with a gating mechanism to enable learning from sparse data, leading to solutions to the assignment problem exhibiting varying degrees of sparsity. Our experimental results on professional soccer tracking data provides insights on learned features and embeddings, as well as on generalization, sensitivity, and network architectural considerations. Our low approximation errors for learned representations and the interactive performance with retrieval times several magnitudes smaller shows that we outperform previous state of the art.

Effectiveness, Explainability and Reliability of Machine Meta-Learning Methods for Predicting Mortality in Patients with COVID-19: Results of the Brazilian COVID-19 Registry

The majority prognostic scores proposed for early assessment of coronavirus disease 19 (COVID-19) patients are bounded by methodological flaws. Our group recently developed a new risk score - ABC2SPH - using traditional statistical methods (least absolute shrinkage and selection operator logistic regression - LASSO). In this article, we provide a thorough comparative study between modern machine learning (ML) methods and state-of-the-art statistical methods, represented by ABC2SPH, in the task of predicting in-hospital mortality in COVID-19 patients using data upon hospital admission. We overcome methodological and technological issues found in previous similar studies, while exploring a large sample (5,032 patients). Additionally, we take advantage of a large and diverse set of methods and investigate the effectiveness of applying meta-learning, more specifically Stacking, in order to combine the methods' strengths and overcome their limitations. In our experiments, our Stacking solutions improved over previous state-of-the-art by more than 26% in predicting death, achieving 87.1% of AUROC and MacroF1 of 73.9%. We also investigated issues related to the interpretability and reliability of the predictions produced by the most effective ML methods. Finally, we discuss the adequacy of AUROC as an evaluation metric for highly imbalanced and skewed datasets commonly found in health-related problems.

Motor Nerve Conduction Block Estimation in Demyelinating Neuropathies by Deconvolution

A deconvolution method is proposed for conduction block (CB) estimation based on two compound muscle action potentials (CMAPs) elicited by stimulating a nerve proximal and distal to the region in which the block is suspected. It estimates the time delay distributions by CMAPs deconvolution, from which CB is computed. The slow afterwave (SAW) is included to describe the motor unit potential, as it gives an important contribution in case of the large temporal dispersion (TD) often found in patients. The method is tested on experimental signals obtained from both healthy subjects and pathological patients, with either Chronic Inflammatory Demyelinating Polyneuropathy (CIDP) or Multifocal Motor Neuropathy (MMN). The new technique outperforms the clinical methods (based on amplitude and area of CMAPs) and a previous state-of-the-art deconvolution approach. It compensates phase cancellations, allowing to discriminate among CB and TD: estimated by the methods of amplitude, area and deconvolution, CB showed a correlation with TD equal to 39.3%, 29.5% and 8.2%, respectively. Moreover, a significant decrease of percentage reconstruction errors of the CMAPs with respect to the previous deconvolution approach is obtained (from a mean/median of 19.1%/16.7% to 11.7%/11.2%). Therefore, the new method is able to discriminate between CB and TD (overcoming the important limitation of clinical approaches) and can approximate patients’ CMAPs better than the previous deconvolution algorithm. Then, it appears to be promising for the diagnosis of demyelinating polyneuropathies, to be further tested in the future in a prospective clinical trial.

DESIGN AND TESTING OF A MULTI-STAGE IP TURBINE FOR FUTURE GEARED TURBOFANS

The multi-stage intermediate pressure turbine (IPT) is a key enabler of the thermodynamic cycle in geared turbofan engine architectures, where fan and turbine rotational speeds become decoupled by employing a power gearbox between them. This allows for the separate aerodynamic optimization of both components, an increase in engine bypass ratios, higher propulsive efficiency, and lower specific fuel consumption. Due to significant aerodynamic differences with conventional low pressure turbines (LPTs), multi-stage IPT designs present new aerodynamic, mechanical and acoustic trade-offs. This work describes the aerodynamic design and experimental validation of a fully featured three-stage IP turbine, including a final row of outlet guide vanes. Experiments have been conducted in a highly engine-representative transonic rotating wind tunnel at the CTA (Centro de Tecnolog'as Aeron'uticas, Spain), in which Mach and Reynolds numbers were matched to engine conditions. The design intent is shown to be fully validated. Efficiency levels are discussed in the context of a previous state-of-the-art LPT, tested in the same facility. It is argued that the efficiency gains of IPTs are due to higher pitch-to-chord ratios, which lead to a reduction in overall profile losses, and higher velocity ratios and lower turning angles, which reduce airfoil secondary flows and three-dimensional losses.

An Enhanced Image Segmentation From 3D to 2D by Using Modified Neural Network

Optical coherence tomography angiography (OCTA) is an imaging which can applied in ophthalmology to provide detailed visualization of the perfusion of vascular networks in the eye. compared to previous state of the art dye-based imaging, such as fluorescein angiography. OCTA is non-invasive, time efficient, and it allows for the examination of retinal vascular in 3D. These advantage of the technique combined with the good usability in commercial devices led to a quick adoption of the new modality in the clinical routine. However, the interpretation of OCTA data is not without problems commonly observed image artifacts and the quite involved algorithmic details of OCTA signal construction can make the clinical assessment of OCTA exams challenging. In this paper we describe the technical background of OCTA and discuss the data acquisition process, common image visualization techniques, as well as 3D to 2D projection using high pass filtering, relu function and convolution neural network (CNN) for more accuracy and segmentation results.

PET Image Reconstruction with Kernel and Kernel Space Composite Regularizer

<div>Represented by the kernelized expectation maximization (KEM), the kernelized maximum-likelihood (ML) expectation maximization (EM) methods have recently gained prominence in PET image reconstruction, outperforming many previous state-of-the-art methods. But they are not immune to the problems of non-kernelized MLEM methods in potentially large reconstruction variance and high sensitivity to iteration number. Also, it is generally difficult to simultaneously reduce image variance and preserve image details using kernels. To solve these problems, this paper presents a novel regularized KEM (RKEM) method with a kernel space composite regularizer for PET image reconstruction. The composite regularizer consists of a convex kernel space graph regularizer that smoothes the kernel coefficients, a non-convex kernel space energy regularizer that enhances the coefficients’ energy, and a composition constant that guarantees the convexity of composite regularizer. These kernel space regularizers are based on the theory of data manifold and graph regularization and can be constructed from different prior image data for simultaneous image variance reduction and image detail preservation. Using this kernel space composite regularizer and the technique of optimization transfer, a globally convergent iterative algorithm is derived for RKEM reconstruction. Tests and comparisons on the simulated and in vivo data are presented to validate and evaluate the proposed algorithm, and demonstrate its better performance and advantages over KEM and other conventional methods.</div>

PET Image Reconstruction with Kernel and Kernel Space Composite Regularizer

<div>Represented by the kernelized expectation maximization (KEM), the kernelized maximum-likelihood (ML) expectation maximization (EM) methods have recently gained prominence in PET image reconstruction, outperforming many previous state-of-the-art methods. But they are not immune to the problems of non-kernelized MLEM methods in potentially large reconstruction variance and high sensitivity to iteration number. Also, it is generally difficult to simultaneously reduce image variance and preserve image details using kernels. To solve these problems, this paper presents a novel regularized KEM (RKEM) method with a kernel space composite regularizer for PET image reconstruction. The composite regularizer consists of a convex kernel space graph regularizer that smoothes the kernel coefficients, a non-convex kernel space energy regularizer that enhances the coefficients’ energy, and a composition constant that guarantees the convexity of composite regularizer. These kernel space regularizers are based on the theory of data manifold and graph regularization and can be constructed from different prior image data for simultaneous image variance reduction and image detail preservation. Using this kernel space composite regularizer and the technique of optimization transfer, a globally convergent iterative algorithm is derived for RKEM reconstruction. Tests and comparisons on the simulated and in vivo data are presented to validate and evaluate the proposed algorithm, and demonstrate its better performance and advantages over KEM and other conventional methods.</div>

Time Series Classification with InceptionFCN

Deep neural networks (DNN) have proven to be efficient in computer vision and data classification with an increasing number of successful applications. Time series classification (TSC) has been one of the challenging problems in data mining in the last decade, and significant research has been proposed with various solutions, including algorithm-based approaches as well as machine and deep learning approaches. This paper focuses on combining the two well-known deep learning techniques, namely the Inception module and the Fully Convolutional Network. The proposed method proved to be more efficient than the previous state-of-the-art InceptionTime method. We tested our model on the univariate TSC benchmark (the UCR/UEA archive), which includes 85 time-series datasets, and proved that our network outperforms the InceptionTime in terms of the training time and overall accuracy on the UCR archive.

iMon

Gaze tracking is a key building block used in many mobile applications including entertainment, personal productivity, accessibility, medical diagnosis, and visual attention monitoring. In this paper, we present iMon, an appearance-based gaze tracking system that is both designed for use on mobile phones and has significantly greater accuracy compared to prior state-of-the-art solutions. iMon achieves this by comprehensively considering the gaze estimation pipeline and then overcoming three different sources of errors. First, instead of assuming that the user's gaze is fixed to a single 2D coordinate, we construct each gaze label using a probabilistic 2D heatmap gaze representation input to overcome errors caused by microsaccade eye motions that cause the exact gaze point to be uncertain. Second, we design an image enhancement model to refine visual details and remove motion blur effects of input eye images. Finally, we apply a calibration scheme to correct for differences between the perceived and actual gaze points caused by individual Kappa angle differences. With all these improvements, iMon achieves a person-independent per-frame tracking error of 1.49 cm (on smartphones) and 1.94 cm (on tablets) when tested with the GazeCapture dataset and 2.01 cm with the TabletGaze dataset. This outperforms the previous state-of-the-art solutions by ~22% to 28%. By averaging multiple per-frame estimations that belong to the same fixation point and applying personal calibration, the tracking error is further reduced to 1.11 cm (smartphones) and 1.59 cm (tablets). Finally, we built implementations that run on an iPhone 12 Pro and show that our mobile implementation of iMon can run at up to 60 frames per second - thus making gaze-based control of applications possible.

Learning discriminative and generalizable features with multi-branch for person re-identification

Finer-grained local features play a supplementary role in the description of pedestrian global features, and the combination of them has been an essential solution to improve discriminative performances in person re-identification (PReID) tasks. The existing part-based methods mostly extract representational semantic parts according to human visual habits or some prior knowledge and focus on spatial partition strategies but ignore the significant influence of channel information on PReID task. So, we proposed an end-to-end multi-branch network architecture (MCSN) jointing multi-level global fusion features, channel features and spatial features in this paper to better learn more diverse and discriminative pedestrian features. It is worth noting that the effect of multi-level fusion features on the performance of the model is taken into account when extracting global features. In addition, to enhance the stability of model training and the generalization ability of the model, the BNNeck and the joint loss function strategy are applied to all vector representation branches. Extensive comparative evaluations are conducted on three mainstream image-based evaluation protocols, including Market-1501, DukeMTMC-ReID and MSMT17, to validate the advantages of our proposed model, which outperforms previous state-of-the-art in ReID tasks.