Unseen Data Research Articles

Insights into the Effects of Tile Size and Tile Overlap Levels on Semantic Segmentation Models Trained for Road Surface Area Extraction from Aerial Orthophotography

Studies addressing the supervised extraction of geospatial elements from aerial imagery with semantic segmentation operations (including road surface areas) commonly feature tile sizes varying from 256 × 256 pixels to 1024 × 1024 pixels with no overlap. Relevant geo-computing works in the field often comment on prediction errors that could be attributed to the effect of tile size (number of pixels or the amount of information in the processed image) or to the overlap levels between adjacent image tiles (caused by the absence of continuity information near the borders). This study provides further insights into the impact of tile overlaps and tile sizes on the performance of deep learning (DL) models trained for road extraction. In this work, three semantic segmentation architectures were trained on data from the SROADEX dataset (orthoimages and their binary road masks) that contains approximately 700 million pixels of the positive “Road” class for the road surface area extraction task. First, a statistical analysis is conducted on the performance metrics achieved on unseen testing data featuring around 18 million pixels of the positive class. The goal of this analysis was to study the difference in mean performance and the main and interaction effects of the fixed factors on the dependent variables. The statistical tests proved that the impact on performance was significant for the main effects and for the two-way interaction between tile size and tile overlap and between tile size and DL architecture, at a level of significance of 0.05. We provide further insights and trends in the predictions of the extensive qualitative analysis carried out with the predictions of the best models at each tile size. The results indicate that training the DL models on larger tile sizes with a small percentage of overlap delivers better road representations and that testing different combinations of model and tile sizes can help achieve a better extraction performance.

DualSyn: A dual-level feature interaction method to predict synergistic drug combinations

Drug combination therapy can reduce drug resistance and improve treatment efficacy, making it an increasingly promising cancer treatment method. Although existing computational methods have achieved significant success, predictions on unseen data remain a challenge. There are complex associations between drug pairs and cell lines, and existing models cannot capture more general feature interaction patterns among them, which hinders the ability of models to generalize from seen samples to unseen samples. To address this problem, we propose a dual-level feature interaction model called DualSyn to efficiently predict the synergy of drug combination therapy. This model first achieves interaction at the drug pair level through the drugs feature extraction module. We also designed two modules to further deepen the interaction at the drug pair and cell line level from two different perspectives. The high-order relation module is used to capture the high-order relationships among the three features, and the global information module focuses on preserving global information details. DualSyn not only improves the AUC by 2.15% compared with the state-of-the-art methods in the transductive task of the benchmark dataset, but also surpasses them in all four tasks under the inductive setting. Overall, DualSyn shows great potential in predicting and explaining drug synergistic therapies, providing a powerful new tool for future clinical applications.

Erosive wear and particle attrition in multi-stage solar particle receivers and screw conveyors: A CFD-DEM approach with machine learning and artificial neural networks

A coupled finite volume and discrete element method (CFD-DEM) numerical model is developed to unravel the fundamental mechanisms of granular transport, surface erosive wear and particle attrition in multistage solar particle receivers (MSPR) and screw conveyors (SC). Additionally, various regression-based supervised machine learning (ML) and artificial neural networks (ANN) are trained and optimized with the goal of quickly quantifying erosion and attrition thereby overcoming (a) the inherent challenges of high computational expense of CFD-DEM simulations and (b) exorbitant time required to post-process and extract DEM results. This study is the first of its kind to harness a synergistic numerical framework (CFD-DEM, ML, ANN) to unravel the fundamental mechanisms of erosion and attrition in MSPRs, non-vibrationally induced SCs, and vibrationally-induced SCs. It is found that the packing distribution, particle curtain width, and severity of particle scattering in a MSPR exhibits a subtle variation with particle diameter. In the context of SCs, the magnitudes of erosion and attrition are contingent on the particle diameter and flow operating conditions such as screw blade velocity, vibrational frequency and vibrational amplitude. Interestingly, a vibrationally-induced SC (VI-SC) exhibits superior performance in terms of low erosion and attrition unlike non-vibrationally induced SC (NVI-SC). It is found that MSPRs with 750–1000 µm particles and SCs with 750–1000 µm coupled with vibrational frequencies and vibrational amplitudes of 5 Hz and 0.007 m, respectively, are the preferred configurations due to minimum particle scattering, low erosive wear and low particle attrition. The trained and optimized ML models such as XGBoost, CatBoost, and Multilayer Perceptron (MLP) exhibit superior performance in predicting erosive wear and particle attrition on unseen DEM data, whereas the Multiple Linear Regression (MLR) and Gaussian Process Regression (GPR) ML exhibit poor performance prediction. For the first time, new information on fundamental granular flow dynamics, erosive wear and particle attrition in MSPRs and SCs is discerned. The methodology allows scientists to quickly examine material degradation for a myriad of industrial applications such as concentrated solar thermal (CST), among others.

DeepIOD: Towards A Context-Aware Indoor-Outdoor Detection Framework Using Smartphone Sensors.

Accurate indoor-outdoor detection (IOD) is essential for location-based services, context-aware computing, and mobile applications, as it enhances service relevance and precision. However, traditional IOD methods, which rely only on GPS data, often fail in indoor environments due to signal obstructions, while IMU data are unreliable on unseen data in real-time applications due to reduced generalizability. This study addresses this research gap by introducing the DeepIOD framework, which leverages IMU sensor data, GPS, and light information to accurately classify environments as indoor or outdoor. The framework preprocesses input data and employs multiple deep neural network models, combining outputs using an adaptive majority voting mechanism to ensure robust and reliable predictions. Experimental results evaluated on six unseen environments using a smartphone demonstrate that DeepIOD achieves significantly higher accuracy than methods using only IMU sensors. Our DeepIOD system achieves a remarkable accuracy rate of 98-99% with a transition time of less than 10 ms. This research concludes that DeepIOD offers a robust and reliable solution for indoor-outdoor classification with high generalizability, highlighting the importance of integrating diverse data sources to improve location-based services and other applications requiring precise environmental context awareness.

The impact of multicentric datasets for the automated tumor delineation in primary prostate cancer using convolutional neural networks on 18F-PSMA-1007 PET

PurposeConvolutional Neural Networks (CNNs) have emerged as transformative tools in the field of radiation oncology, significantly advancing the precision of contouring practices. However, the adaptability of these algorithms across diverse scanners, institutions, and imaging protocols remains a considerable obstacle. This study aims to investigate the effects of incorporating institution-specific datasets into the training regimen of CNNs to assess their generalization ability in real-world clinical environments. Focusing on a data-centric analysis, the influence of varying multi- and single center training approaches on algorithm performance is conducted.MethodsnnU-Net is trained using a dataset comprising 161 18F-PSMA-1007 PET images collected from four distinct institutions (Freiburg: n = 96, Munich: n = 19, Cyprus: n = 32, Dresden: n = 14). The dataset is partitioned such that data from each center are systematically excluded from training and used solely for testing to assess the model's generalizability and adaptability to data from unfamiliar sources. Performance is compared through a 5-Fold Cross-Validation, providing a detailed comparison between models trained on datasets from single centers to those trained on aggregated multi-center datasets. Dice Similarity Score, Hausdorff distance and volumetric analysis are used as primary evaluation metrics.ResultsThe mixed training approach yielded a median DSC of 0.76 (IQR: 0.64–0.84) in a five-fold cross-validation, showing no significant differences (p = 0.18) compared to models trained with data exclusion from each center, which performed with a median DSC of 0.74 (IQR: 0.56–0.86). Significant performance improvements regarding multi-center training were observed for the Dresden cohort (multi-center median DSC 0.71, IQR: 0.58–0.80 vs. single-center 0.68, IQR: 0.50–0.80, p < 0.001) and Cyprus cohort (multi-center 0.74, IQR: 0.62–0.83 vs. single-center 0.72, IQR: 0.54–0.82, p < 0.01). While Munich and Freiburg also showed performance improvements with multi-center training, results showed no statistical significance (Munich: multi-center DSC 0.74, IQR: 0.60–0.80 vs. single-center 0.72, IQR: 0.59–0.82, p > 0.05; Freiburg: multi-center 0.78, IQR: 0.53–0.87 vs. single-center 0.71, IQR: 0.53–0.83, p = 0.23).ConclusionCNNs trained for auto contouring intraprostatic GTV in 18F-PSMA-1007 PET on a diverse dataset from multiple centers mostly generalize well to unseen data from other centers. Training on a multicentric dataset can improve performance compared to training exclusively with a single-center dataset regarding intraprostatic 18F-PSMA-1007 PET GTV segmentation. The segmentation performance of the same CNN can vary depending on the dataset employed for training and testing.

An improved scheme for passive cavitation imaging in high intensity focused ultrasound

Passive cavitation imaging (PCI) is an effective technique for monitoring high intensity focused ultrasound (HIFU) procedure due to the correlation between cavitation activities and the therapeutic effects. Nevertheless, current PCI techniques are unable to recognize noise signals from non-cavitation situations, and they fail to accurately visualize the highly dynamic cavitation activities due to the lack of a stable baseline reference. In this study, a deep learning-based cavitation detection network (CD-NET) was proposed to determine whether passively received channel signals originated from cavitation activities or were merely noises. Additionally, a cavitation intensity baseline was derived to better display the drastically time-varying cavitation activities. The CD-NET model was trained and validated with data collected from in-vitro and ex-vivo HIFU sonication experiments, and then tested by in-vivo rabbit experiments. Both CD-NET models, trained with data from the in-vitro agar phantoms and ex-vivo bovine liver tissues, successfully distinguished between cavitation and non-cavitation activities from the same study object, achieving accuracies as high as 99.7% and 99.5%, respectively. In addition, the models trained on in-vitro and ex-vivo datasets achieved accuracies of 88.1% and 91.4%, respectively, when detecting cavitation events from unseen in-vivo data. The results demonstrate that the improved PCI scheme, utilizing the CD-NET model and a data-derived cavitation baseline, is resistant to noise from non-cavitation situations and can depict the dynamic characteristics of cavitation more reliably. In conclusion, this study addressed two practical challenges in PCI techniques, allowing for better monitoring of cavitation-related HIFU sonication in clinical applications.

Deformation stage identification in steel material using acoustic emission with a hybrid denoising method and artificial neural network

Acoustic emission (AE) is widely used for identifying source mechanisms and the deformation stage of steel material. The effectiveness of this non-destructive monitoring technique heavily depends on the quality of the measured AE signals. However, the AE signals from deformation are easily contaminated by the signals from noise in a noisy environment. This paper presents a hybrid model for deformation stage identification, which combines a self-adaptive denoising technique and an Artificial neural network (ANN). In pursuit of model generality, AE signals were collected from tensile coupon tests with various steel materials and loading speeds. First, a decomposition-based denoising method is applied based on the singular spectral analysis (SSA) and variational mode decomposition (VMD), which is defined as SSA-VMD. Its effectiveness is demonstrated by simulated signals and experimental results. Following the use of the denoising technique, an ANN is constructed to identify the deformation stage of steel materials with the input of features extracted from the filtered AE signals. The results indicate that the ANN achieves a high prediction accuracy of 0.93 in the test set and 0.87 in unseen data. By applying this denoising method, the ANN-based approach enables accurate correlation of the collected AE signals to deformation stages. The finding can be used as the basis for the creation of new methodologies for monitoring structural health status of in-service steel structures.

Efficient predictive modeling of resilient modulus in stabilized clayey soil using automated machine learning

Pavement subgrade design relies on the resilient modulus (Mr) to analyze structural response to vehicle-like loading. Adding stabilizers to the soil subgrade makes estimating Mr difficult and resource-intensive. This study uses an automated machine learning (ML) strategy to predict the Mr of stabilized clayey soil using recycled plastic waste. The proposed method automates model selection and hyperparameter tuning, making it a feasible alternative to tedious ML modeling and costly laboratory testing. From extensive laboratory investigation involving 3285 experimental data points, the automated ML model using Bayesian optimization evaluates ensembles, support vector machine (SVM), neural network (NET), decision trees, (TREE), and Gaussian process (GP) regression models, identifying the best model based on cross-validation mean squared error (MSE). Bayesian optimization explores hyperparameter spaces to find optimal configurations, enhancing the accuracy, scalability, and reliability of the prediction model. The optimization process yielded the best results for the ensemble least square boost (LSBoost) model with a cross-validation mean squared error (MSE) value of 6.723×10−29. The optimized ML model’s performance is measured using R2 and adjusted R2. The LSboost model’s R2 and adjusted R2 values of 0.9999 suggested overfitting, prompting further investigations using performance metrics like root mean squared error (RMSE), mean absolute error (MSE), and probability density function (PDF) for normalized absolute error (NAE) for training and testing datasets for the predictive ML model. The small RMSE and MAE (0.0049 and 0.0005) values and symmetrical NAE distribution of the proposed ML model demonstrate its high accuracy and generalization capabilities. The proposed model was subsequently tested on the new, unseen data and achieved predictions with an error rate of 0.24 %. This confirms the proposed ML model’s superiority over conventional deformation models, making it ideal for reliable geotechnical engineering applications.

Deep Learning Forecasts Caldera Collapse Events at Kı̄lauea Volcano

AbstractDuring the 3 month long eruption of Kı̄lauea volcano, Hawaii in 2018, the pre‐existing summit caldera collapsed in over 60 quasi‐periodic failure events. The last 40 of these events, which generated Mw &gt; 5 very long period (VLP) earthquakes, had inter‐event times between 0.8 and 2.2 days. These failure events offer a unique data set for testing methods for predicting earthquake recurrence based on locally recorded GPS, tilt, and seismicity data. In this work, we train a deep learning graph neural network (GNN) to predict the time‐to‐failure of the caldera collapse events using only a fraction of the data recorded at the start of each cycle. We find that the GNN generalizes to unseen data and can predict the time‐to‐failure to within a few hours using only 0.5 days of data, substantially improving upon a null model based only on inter‐event statistics. Predictions improve with increasing input data length, and are most accurate when using high‐SNR tilt‐meter data. Applying the trained GNN to synthetic data with different magma‐chamber pressure decay times predicts failure at a nearly constant stress threshold, revealing that the GNN is sensing the underling physics of caldera collapse. These findings demonstrate the predictability of caldera collapse sequences under well monitored conditions, and highlight the potential of machine learning methods for forecasting real world catastrophic events with limited training data.

Assessing screw length impact on bone strain in proximal humerus fracture fixation via surrogate modelling.

A high failure rate is associated with fracture plates in proximal humerus fractures. The causes of failure remain unclear due to the complexity of the problem including the number and position of the screws, their length and orientation in the space. Finite element (FE) analysis has been used for the analysis of plating of proximal humeral fractures, but due to computational costs is unable to fully explore all potential screw combinations. Surrogate modelling is a viable solution, having the potential to significantly reduce the computational cost whilst requiring a moderate number of training sets. This study aimed to develop adaptive neural network (ANN)-based surrogate models to predict the strain in the humeral bone as a result of changing the length of the screws. The ANN models were trained using data from FE simulations of a single humerus, and after defining the best training sample size, multiple and single-output models were developed. The best performing ANN model was used to predict all the possible screw length configurations. The ANN predictions were compared with the FE results of unseen data, showing a good correlation (R2 = 0.99) and low levels of error (RMSE = 0.51%-1.83% strain). The ANN predictions of all possible screw length configurations showed that the screw that provided the medial support was the most influential on the predicted strain. Overall, the ANN-based surrogate model accurately captured bone strains and has the potential to be used for more complex problems with a larger number of variables.

Predictive analytics for financial compliance: Machine learning concepts for fraudulent transaction identification

Predictive analytics has emerged as a pivotal tool in financial compliance, offering sophisticated methods for identifying fraudulent transactions through the application of machine learning (ML) concepts. As financial institutions grapple with increasingly complex fraud schemes and stringent regulatory requirements, the integration of predictive analytics with ML provides a proactive approach to fraud detection and prevention. Machine learning algorithms excel in analyzing vast datasets, identifying hidden patterns, and making real-time predictions. In the realm of financial compliance, supervised learning models such as logistic regression, decision trees, and random forests are commonly used to classify transactions as legitimate or fraudulent. These models are trained on historical transaction data, learning to recognize the subtle indicators of fraud by identifying correlations between various features and fraudulent outcomes. This allows for high-accuracy predictions on new, unseen data. Unsupervised learning techniques, such as clustering and anomaly detection, are equally critical in predictive analytics for financial compliance. These methods do not require labeled data and are adept at uncovering novel fraud patterns by detecting outliers and irregularities that deviate from normal transactional behavior. Anomaly detection algorithms, including k-means clustering and isolation forests, can identify transactions that exhibit unusual characteristics, flagging them for further investigation. The integration of predictive analytics with real-time data processing capabilities enhances the agility of fraud detection systems. Streaming analytics and real-time scoring enable the continuous monitoring of transactions, ensuring that suspicious activities are identified and addressed promptly. This real-time aspect is crucial for minimizing the impact of fraudulent transactions and ensuring compliance with regulatory standards. Despite the advancements, implementing predictive analytics for financial compliance involves challenges such as ensuring data quality, addressing privacy concerns, and maintaining model transparency. Financial institutions must navigate these challenges by employing robust data governance practices, leveraging secure data processing techniques, and adopting explainable AI models that provide insights into their decision-making processes. In conclusion, predictive analytics, powered by machine learning concepts, offers a robust framework for identifying fraudulent transactions and enhancing financial compliance. By leveraging advanced ML algorithms and real-time data processing, financial institutions can proactively detect and prevent fraud, thereby safeguarding their operations and ensuring adherence to regulatory mandates. This approach not only mitigates financial losses but also strengthens the overall integrity and trustworthiness of the financial system.

Exploring the impact of thiol collectors system on copper sulfide flotation through machine learning-driven modeling

Collector selection is a critical step in flotation, as it has a direct impact on product quality, flotation recovery, and selectivity. Collectors can consist of different components, and their effectiveness can vary depending on the type of ore being processed. The general practice in both literature and in industry is to use a mixture of collectors rather than a single collector. However, the use of a collector mixture introduces several complex issues. It is challenging to determine the specific effects of each collector on different minerals, as well as to understand the synergistic effects of mixed collectors in flotation. This study presents a novel investigation focusing on the impact of blends of NAX, AEROPHINE® 3422, and AERO® MX 5149, in varying dosages and combinations, on the flotation performance of Kupferschiefer copper ore. Kinetics flotation tests were conducted using a mechanical flotation cell with various combinations and dosages of listed collectors. For this investigation, different predictive models such as machine-learning (ML) and conventional regression analyses were developed. For model construction, a database including the results of comprehensive experimental results was constructed. The best performing model was selected considering statistical performance indicators and their performance on unseen data. A sensitivity analysis was conducted on the model to justify contributions of collectors on the copper recovery and grade. The results showed that the ML-based models provide compatible results with the expert opinions and have higher statistical performance than conventional modelling tools. According to the experimental results and models’ findings, it has shown that AEROPHINE® 3422 (a blend of isopropyl ethyl thionocarbamate and dithiophosphinate) was the most influential collector for the copper recovery. In addition, two ternary graphs were generated from the modeled data to formulate mixtures for different grades and recovery priorities.

Holistic evaluation of a machine learning-based timing calibration for PET detectors under varying data sparsity

Objective. Modern PET scanners offer precise TOF information, improving the SNR of the reconstructed images. Timing calibrations are performed to reduce the worsening effects of the system components and provide valuable TOF information. Traditional calibration procedures often provide static or linear corrections, with the drawback that higher-order skews or event-to-event corrections are not addressed. Novel research demonstrated significant improvements in the reachable timing resolutions when combining conventional calibration approaches with machine learning, with the disadvantage of extensive calibration times infeasible for a clinical application. In this work, we made the first steps towards an in-system application and analyzed the effects of varying data sparsity on a machine learning timing calibration, aiming to accelerate the calibration time. Furthermore, we demonstrated the versatility of our calibration concept by applying the procedure for the first time to analog readout technology. Approach. We modified experimentally acquired calibration data used for training regarding their statistical and spatial sparsity, mimicking reduced measurement time and variability of the training data. Trained models were tested on unseen test data, characterized by fine spatial sampling and rich statistics. In total, 80 decision tree models with the same hyperparameter settings, were trained and holistically evaluated regarding data scientific, physics-based, and PET-based quality criteria. Main results. The calibration procedure can be heavily reduced from several days to some minutes without sacrificing quality and still significantly improving the timing resolution from (304±5)ps to (216±1)ps compared to conventionally used analytical calibration methods.Significance. This work serves as the first step in making the developed machine learning-based calibration suitable for an in-system application to profit from the method’s capabilities on the system level. Furthermore, this work demonstrates the functionality of the methodology on detectors using analog readout technology. The proposed holistic evaluation criteria here serve as a guideline for future evaluations of machine learning-based calibration approaches.

EEG-DG: A Multi-Source Domain Generalization Framework for Motor Imagery EEG Classification.

Motor imagery EEG classification plays a crucial role in non-invasive Brain-Computer Interface (BCI) research. However, the performance of classification is affected by the non-stationarity and individual variations of EEG signals. Simply pooling EEG data with different statistical distributions to train a classification model can severely degrade the generalization performance. To address this issue, the existing methods primarily focus on domain adaptation, which requires access to the test data during training. This is unrealistic and impractical in many EEG application scenarios. In this paper, we propose a novel multi-source domain generalization framework called EEG-DG, which leverages multiple source domains with different statistical distributions to build generalizable models on unseen target EEG data. We optimize both the marginal and conditional distributions to ensure the stability of the joint distribution across source domains and extend it to a multi-source domain generalization framework to achieve domain-invariant feature representation, thereby alleviating calibration efforts. Systematic experiments conducted on a simulative dataset, BCI competition IV 2a, 2b, and OpenBMI datasets, demonstrate the superiority and competitive performance of our proposed framework over other state-of-the-art methods. Specifically, EEG-DG achieves average classification accuracies of 81.79% and 87.12% on datasets IV-2a and IV-2b, respectively, and 78.37% and 76.94% for inter-session and inter-subject evaluations on dataset OpenBMI, which even outperforms some domain adaptation methods. Our code is available at https://github.com/zxchit2022/EEG-DG for evaluation.

Deep learning-based prediction of 3-dimensional silver contact shapes enabling improved quality control in solar cell metallization

The industrial metallization of Si solar cells predominantly relies on screen printing, with silver as the preferred electrode material. However, the design of commercial screens often leads to suboptimal silver usage and increased electrical resistance due to print-related inhomogeneities like mesh marks, constrictions and spreading. Real-time monitoring of quality parameters during production has thus become increasingly critical. Current inline optical quality control systems usually only include 2D visualizations of the printed layout, which limits their effectiveness in quality control. Options that allow 3D measurements are usually slow, expensive, and therefore not worth considering in most cases. This research focuses on the development of a model that can estimate the three-dimensional shape of printed contact fingers from a single 2D image without the need of additional hardware using deep learning. Furthermore, a workflow for the generation of training data, which involves the creation of image pairs from a 2D microscope and a 3D confocal laser scanning microscope (CLSM) to accurately represent solar cell fingers, is presented. After model training, the predicted height maps are compared with the ground truth height maps, and the robustness of the model with respect to a paste variation and screen parameter variation is examined. The results confirm the feasibility and reliability of deep learning-based 3D shape estimation, extending its applicability to new, previously unseen data from screen-printed contact fingers. With a structural similarity index (SSIM) score of 0.76, a strong correlation between the estimated and ground truth height maps is established. In summary, our deep learning-based approach for height map estimation offers an effective and reliable solution for fast inline detection and analysis of the cross-sectional area of the printed contact fingers.

Φ-DVAE: Physics-informed dynamical variational autoencoders for unstructured data assimilation

Incorporating unstructured data into physical models is a challenging problem that is emerging in data assimilation. Traditional approaches focus on well-defined observation operators whose functional forms are typically assumed to be known. This prevents these methods from achieving a consistent model-data synthesis in configurations where the mapping from data-space to model-space is unknown. To address these shortcomings, in this paper we develop a physics-informed dynamical variational autoencoder (Φ-DVAE) to embed diverse data streams into time-evolving physical systems described by differential equations. Our approach combines a standard, possibly nonlinear, filter for the latent state-space model and a VAE, to assimilate the unstructured data into the latent dynamical system. Unstructured data, in our example systems, comes in the form of video data and velocity field measurements, however the methodology is suitably generic to allow for arbitrary unknown observation operators. A variational Bayesian framework is used for the joint estimation of the encoding, latent states, and unknown system parameters. To demonstrate the method, we provide case studies with the Lorenz-63 ordinary differential equation, and the advection and Korteweg-de Vries partial differential equations. Our results, with synthetic data, show that Φ-DVAE provides a data efficient dynamics encoding methodology which is competitive with standard approaches. Unknown parameters are recovered with uncertainty quantification, and unseen data are accurately predicted.

Interpretable deep residual network uncovers nucleosome positioning and associated features.

Nucleosomes represent elementary building units of eukaryotic chromosomes and consist of DNA wrapped around a histone octamer flanked by linker DNA segments. Nucleosomes are central in epigenetic pathways and their genomic positioning is associated with regulation of gene expression, DNA replication, DNA methylation and DNA repair, among other functions. Building on prior discoveries that DNA sequences noticeably affect nucleosome positioning, our objective is to identify nucleosome positions and related features across entire genome. Here, we introduce an interpretable framework based on the concepts of deep residual networks (NuPoSe). Trained on high-coverage human experimental MNase-seq data, NuPoSe is able to learn sequence and structural patterns associated with nucleosome organization in human genome. NuPoSe can bealsoapplied to unseen data from different organisms and cell types. Our findings point to 43 informative features, most of them constitute tri-nucleotides, di-nucleotides and one tetra-nucleotide. Most features are significantly associated with the nucleosomal structuralcharacteristics, namely, periodicity of nucleosomal DNA and its location with respect to a histone octamer. Importantly, we show that features derived from the 27 bp linker DNAflanking nucleosomescontribute up to 10% to the quality of the prediction model. This,alongwith the comprehensive training sets, deep-learning architecture, and feature selection method,may contribute totheNuPoSe's 80-89% classificationaccuracy on different independentdatasets.

MOB-Net: Limb-modularized uncertainty torque learning of humanoids for sensorless external torque estimation

Momentum observer (MOB) can estimate external joint torque without requiring additional sensors, such as force/torque or joint torque sensors. However, the estimation performance of MOB deteriorates due to the model uncertainty which encompasses the modeling errors and the joint friction. Moreover, the estimation error is significant when MOB is applied to high-dimensional floating-base humanoids, which prevents the estimated external joint torque from being used for force control or collision detection in the real humanoid robot. In this paper, the pure external joint torque estimation method named MOB-Net, is proposed for humanoids. MOB-Net learns the model uncertainty torque and calibrates the estimated signal of MOB, substantially reducing the estimation errors of MOB. The external joint torque can be estimated in the generalized coordinate including whole-body and virtual joints of the floating-base robot with only internal sensors (an IMU on the pelvis and encoders in the joints). Furthermore, MOB-Net shows more robust performance for the unseen data compared to the end-to-end learning method, and the robustness of MOB-Net is validated through extensive simulations, real robot experiments, and ablation studies. Finally, various collision handling scenarios are presented to show the versatility of MOB-Net: contact wrench feedback control for locomotion, collision detection, and collision reaction for safety.

Deep learning-based image quality assessment for optical coherence tomography macular scans: a multicentre study

AimsTo develop and externally test deep learning (DL) models for assessing the image quality of three-dimensional (3D) macular scans from Cirrus and Spectralis optical coherence tomography devices.MethodsWe retrospectively collected two...

FetMRQC: A robust quality control system for multi-centric fetal brain MRI

Fetal brain MRI is becoming an increasingly relevant complement to neurosonography for perinatal diagnosis, allowing fundamental insights into fetal brain development throughout gestation. However, uncontrolled fetal motion and heterogeneity in acquisition protocols lead to data of variable quality, potentially biasing the outcome of subsequent studies. We present FetMRQC, an open-source machine-learning framework for automated image quality assessment and quality control that is robust to domain shifts induced by the heterogeneity of clinical data. FetMRQC extracts an ensemble of quality metrics from unprocessed anatomical MRI and combines them to predict experts’ ratings using random forests. We validate our framework on a pioneeringly large and diverse dataset of more than 1600 manually rated fetal brain T2-weighted images from four clinical centers and 13 different scanners. Our study shows that FetMRQC’s predictions generalize well to unseen data while being interpretable. FetMRQC is a step towards more robust fetal brain neuroimaging, which has the potential to shed new insights on the developing human brain.