Deep Learning Methods Research Articles

Metastasis Detection Using True and Artificial T1-Weighted Postcontrast Images in Brain MRI

Objectives Small lesions are the limiting factor for reducing gadolinium-based contrast agents in brain magnetic resonance imaging (MRI). The purpose of this study was to compare the sensitivity and precision in metastasis detection on true contrast-enhanced T1-weighted (T1w) images and artificial images synthesized by a deep learning method using low-dose images. Materials and Methods In this prospective, multicenter study (5 centers, 12 scanners), 917 participants underwent brain MRI between October 2021 and March 2023 including T1w low-dose (0.033 mmol/kg) and full-dose (0.1 mmol/kg) images. Forty participants with metastases or unremarkable brain findings were evaluated in a reading (mean age ± SD, 54.3 ± 15.1 years; 24 men). True and artificial T1w images were assessed for metastases in random order with 4 weeks between readings by 2 neuroradiologists. A reference reader reviewed all data to confirm metastases. Performances were compared using mid-P McNemar tests for sensitivity and Wilcoxon signed rank tests for false-positive findings. Results The reference reader identified 97 metastases. The sensitivity of reader 1 did not differ significantly between sequences (sensitivity [precision]: true, 66.0% [98.5%]; artificial, 61.9% [98.4%]; P = 0.38). With a lower precision than reader 1, reader 2 found significantly more metastases using true images (sensitivity [precision]: true, 78.4% [87.4%]; artificial, 60.8% [80.8%]; P < 0.001). There was no significant difference in sensitivity for metastases ≥5 mm. The number of false-positive findings did not differ significantly between sequences. Conclusions One reader showed a significantly higher overall sensitivity using true images. The similar detection performance for metastases ≥5 mm is promising for applying low-dose imaging in less challenging diagnostic tasks than metastasis detection.

Prediction of Electrical Characteristics of Fin Field-effect Transistor Devices Based on Simulation Using Deep Learning Method

Prediction of Electrical Characteristics of Fin Field-effect Transistor Devices Based on Simulation Using Deep Learning Method

Real-time disease detection on bean leaves from a small image dataset using data augmentation and deep learning methods

Real-time disease detection on bean leaves from a small image dataset using data augmentation and deep learning methods

Objective classification for solid hydrometeor particles using deep learning

AbstractVarious small particles are present in the clouds. Hydrometeor videosonde (HYVIS) is an in situ instrument to observe such particles. The movies were used to capture hydrometeors with approximately 110,000 frames per sounding. When the particles in such movies were manually classified and measured sizes of particles for every several frames, it took an unrealistically long time to perform statistical analysis because of the large number of observed frames. Particle classification is subjective for the observers. This study developed a technique to classify cloud particles objectively using deep learning to overcome these problems and investigated the statistical microphysical characteristics of clouds. This study used the deep learning method You Only Looking Once for detection and classification. The training data were obtained using HYVIS in the Republic of Palau in 2013. The results trained using only HYVIS images showed low validity because some types of particles in the training data were insufficient. The data for typical particle shapes were augmented to improve the classification. Thus, the validity of classification using augmented data was improved. We used the Yonaguni Island HYVIS observation results from manual and artificial intelligence (AI) classification. The AI tended to classify the less irregular type than the manual, but no other significant differences were found. We believe this AI classification system will be for cloud microphysical studies on solid-phase particles.

Masked Graph Autoencoder for Self‐Supervised Transportation Mode Recognition

ABSTRACTTransportation mode recognition reveals the traveling patterns of urban residents and plays an important role in travel analysis and city planning. Research has been done to identify transportation modes using deep learning methods. However, large‐scale and far‐ranging labeled trajectory data are hard to acquire, which hinders the training and application of supervised models. In this study, we propose a self‐supervised method for trajectory transportation modes recognition. First, we construct sequence and dependency graph for trajectory points. Next, we apply mask and remask strategy to graph node features to enhance its learning ability in semantic information. Notably, in order to learn features without labeled data, we introduce graph autoencoder into the framework. This method is able to learn the sequence and dependency feature reconstruction of the travel paths. It can obtain underlying semantic information with the mask and remask strategy. We applied our model to GPS trajectory dataset from the Microsoft GeoLife Project, achieving an average accuracy of 68.26% in identifying the transportation modes among seven categories. The result indicates that the proposed method offers a solution to identifying transportation modes without abundant labeled data. This method can also help to extract space–time features at different scales from trajectory data, which could be applied to various downstream tasks.

Spatial Prediction of Soil Continuous and Categorical Properties Using Deep Learning Approaches for Tamil Nadu, India

Large-scale mapping of soil resources can be crucial and indispensable for several of the managerial applications and policy implications. With machine learning models being the most utilized modeling technique for digital soil mapping (DSM), the implementation of model-based deep learning methods for spatial soil predictions is still under scrutiny. In this study, soil continuous (pH and OC) and categorical variables (order and suborder) were predicted using deep learning–multi layer perceptron (DL-MLP) and one-dimensional convolutional neural networks (1D-CNN) for the entire state of Tamil Nadu, India. For training the deep learning models, 27,098 profile observations (0–30 cm) were extracted from the generated soil database, considering soil series as the distinctive stratum. A total of 43 SCORPAN-based environmental covariates were considered, of which 37 covariates were retained after the recursive feature elimination (RFE) process. The validation and test results obtained for each of the soil attributes for both the algorithms were most comparable with the DL-MLP algorithm depicting the attributes’ most intricate spatial organization details, compared to the 1D-CNN model. Irrespective of the algorithms and datasets, the R2 and RMSE values of the pH attribute ranged from 0.15 to 0.30 and 0.97 to 1.15, respectively. Similarly, the R2 and RMSE of the OC attribute ranged from 0.20 to 0.39 and 0.38 to 0.42, respectively. Further, the overall accuracy (OA) of the order and suborder classification ranged from 39% to 67% and 35% to 64%, respectively. The explicit quantification of the covariate importance derived from the permutation feature importance implied that both the models tried to incorporate the covariate importance with respect to the genesis of the soil attribute under study. Such approaches of the deep learning models integrating soil–environmental relationships under limited parameterization and computing costs can serve as a baseline study, emphasizing opportunities in increasing the transferability and generalizability of the model while accounting for the associated environmental dependencies.

Application of quantitative histomorphometric features in computational pathology

AbstractComputer vision has facilitated the execution of various computer‐aided diagnostic tasks. From a methodological perspective, these tasks are primarily implemented using two dominant strategies: end‐to‐end Deep learning (DL)‐based methods and traditional feature engineering‐based methods. DL methods are capable of automatically extracting, analyzing, and filtering features, leading to final decision‐making from whole slide images. However, these methods are often criticized for the “black box” issue, a significant limitation of DL. In contrast, traditional feature engineering‐based methods involve well‐defined quantitative input features. But it was considered as less potent than DL methods. Advances in segmentation technology and the development of quantitative histomorphometric (QH) feature representation have propelled the evolution of feature engineering‐based methods. This review contrasts the performance differences between the two methods and traces the development of QH feature representation. The conclusion is that, with the ongoing progress in QH feature representation and segmentation technology, methods based on QH features will leverage their advantages—such as explainability, reduced reliance on large training datasets, and lower computational resource requirements—to play a more significant role in some clinical tasks. They may even replace DL methods somewhat or be used in conjunction with them to achieve accurate and understandable results.

Multi-type stroke lesion segmentation: comparison of single-stage and hierarchical approach.

Stroke, a major cause of death and disability worldwide, can be haemorrhagic or ischaemic depending on the type of bleeding in the brain. Rapid and accurate identification of stroke type and lesion segmentation is critical for timely and effective treatment. However, existing research primarily focuses on segmenting a single stroke type, potentially limiting their clinical applicability. This study addresses this gap by exploring multi-type stroke lesion segmentation using deep learning methods. Specifically, we investigate two distinct approaches: a single-stage approach that directly segments all tissue types in one model and a hierarchical approach that first classifies stroke types and then utilises specialised segmentation models for each subtype. Recognising the importance of accurate stroke classification for the hierarchical approach, we evaluate ResNet, ResNeXt and ViT networks, incorporating focal loss and oversampling techniques to mitigate the impact of class imbalance. We further explore the performance of U-Net, U-Net++ and DeepLabV3 models for segmentation within each approach. We use a comprehensive dataset of 6650 images provided by the Ministry of Health of the Republic of Türkiye. This dataset includes 1130 ischaemic strokes, 1093 haemorrhagic strokes and 4427 non-stroke cases. In our comparative experiments, we achieve an AUC score of 0.996 when classifying stroke and non-stroke slices. For lesion segmentation task, while the performance of different architectures is comparable, the hierarchical training approach outperforms the single-stage approach in terms of intersection over union (IoU). The performance of the U-Net model increased significantly from an IoU of 0.788 to 0.875 when the hierarchical approach is used. This comparative analysis aims to identify the most effective approach and deep learning model for multi-type stroke lesion segmentation in brain CT scans, potentially leading to improved clinical decision-making, treatment efficiency and outcomes.

An efficient deep learning method for amino acid substitution model selection.

Amino acid substitution models play an important role in studying the evolutionary relationships among species from protein sequences. The amino acid substitution model consists of a large number of parameters; therefore, it is estimated from hundreds or thousands of alignments. Both general models and clade-specific models have been estimated and widely used in phylogenetic analyses. The maximum likelihood method is normally used to select the best fit model for a specific protein alignment under the study. A number of studies have discussed theoretical concerns as well as computational burden of the maximum likelihood methods in model selection. Recently, machine learning methods have been proposed for selecting nucleotide models. In this paper, we propose a method to measure substitution rates among amino acids (called summary statistics) from protein alignments to efficiently train a deep learning network of so-called ModelDetector for detecting amino acid substitution models. The ModelDetector network was trained from 2,246,400 alignments on a computer with 8 cores (without GPU) in about 3.3 hours. Experiments on simulation data showed that the accuracy of the ModelDetector was comparable with that of the maximum likelihood method ModelFinder. It was orders of magnitude faster than the maximum likelihood method in inferring amino acid substitution models and able to analyze genome alignments with millions of sites in minutes. The results indicate that the deep learning network can play as a promising tool for amino acid substitution model selection.

Hyperspectral Imaging Combined with Deep Learning for the Early Detection of Strawberry Leaf Gray Mold Disease

The presence of gray mold can seriously affect the yield and quality of strawberries. Due to their susceptibility and the rapid spread of this disease, it is important to develop early, accurate, rapid, and non-destructive disease identification strategies. In this study, the early detection of strawberry leaf diseases was performed using hyperspectral imaging combining multi-dimensional features like spectral fingerprints and vegetation indices. Firstly, hyperspectral images of healthy and early affected leaves (24 h) were acquired using a hyperspectral imaging system. Then, spectral reflectance (616) and vegetation index (40) were extracted. Next, the CARS algorithm was used to extract spectral fingerprint features (17). Pearson correlation analysis combined with the SPA method was used to select five significant vegetation indices. Finally, we used five deep learning methods (LSTMs, CNNs, BPFs, and KNNs) to build disease detection models for strawberries based on individual and fusion characteristics. The results showed that the accuracy of the recognition model based on fused features ranged from 88.9% to 96.6%. The CNN recognition model based on fused features performed best, with a recognition accuracy of 96.6%. Overall, the fused feature-based model can reduce the dimensionality of the classification data and effectively improve the predicting accuracy and precision of the classification algorithm.

Deep active learning for multi label text classification.

Given a set of labels, multi-label text classification (MLTC) aims to assign multiple relevant labels for a text. Recently, deep learning models get inspiring results in MLTC. Training a high-quality deep MLTC model typically demands large-scale labeled data. And comparing with annotations for single-label data samples, annotations for multi-label samples are typically more time-consuming and expensive. Active learning can enable a classification model to achieve optimal prediction performance using fewer labeled samples. Although active learning has been considered for deep learning models, there are few studies on active learning for deep multi-label classification models. In this work, for the deep MLTC model, we propose a deep Active Learning method based on Bayesian deep learning and Expected confidence (BEAL). It adopts Bayesian deep learning to derive the deep model's posterior predictive distribution and defines a new expected confidence-based acquisition function to select uncertain samples for deep MLTC model training. Moreover, we perform experiments with a BERT-based MLTC model, where BERT can achieve satisfactory performance by fine-tuning in various classification tasks. The results on benchmark datasets demonstrate that BEAL enables more efficient model training, allowing the deep model to achieve training convergence with fewer labeled samples.

Intelligent Prestack Multi-trace Seismic Inversion Constrained by Probabilistic Geological Information

Seismic inversion uses observed data including well-logging and seismic data to infer rock parameters. However, it still suffers from non-unique solutions. The non-uniqueness increase when extrapolating away from well location. To mitigate this effect, we incorporate geological information into a deep learning (DL) framework in the form of generated probabilistic labels. We transform geological information into reservoir heterogeneity (RH) weights and geological pattern constraint (GPC) weights to generate probabilistic labels. The RH weights are used to capture the variability in accuracy during the spatial extrapolation of well data. The RH weights are used to balance the constraint weights between well and seismic data in the DL model. The GPC weights are used to characterize the changes of geological patterns from well locations to non-well locations. Based on variations in local prestack seismic waveforms, the DL model learns extrapolation patterns of well data with similar geological patterns. By combining the rock parameters derived from well logs with these two types of weights, we can generate a series of probability labels. Two types of geological information and well-logging data are deeply integrated into a supervised 2D-ResUNet framework. In this way, we develop a novel DL-based prestack multi-trace seismic inversion method. We validate our method on a 3D synthetic model and a 3D field dataset. The results show that our method exhibits great potential for characterizing the spatial variation of subsurface rocks and outperforms traditional deep learning methods.

Optimising bearing degradation prediction: a bayesian vector autoregressive-assisted DeepAR model approach

Abstract Accurate prediction of the remaining useful life of bearings is crucial for the maintenance of rotating machinery. Current research often involves extracting multidimensional features from vibration signals to construct health degradation indicator, and health indicator is directly used for degradation prediction. However, constructed health indicator exhibit complex nonlinearity and fail to utilize the correlations with the original multidimensional features. To address this limitation, we propose an integrated prediction framework, known as the bayesian vector autoregression-assisted DeepAR method. The proposed two step prediction framework combines the advantages of covariance prediction and temporal correlation prediction, and realizes the full utilization of multidimensional features correlation information. Additionally, the proposed approach utilizes the intrinsic properties of statistical models to constrain improve the prediction results of deep learning methods, achieving nonlinear modeling of the degradation indicator while maintaining the stability of the prediction results. This process of prediction, re-estimation and adaptive adjustment achieves a balance in model performance. Finally, experimental studies using the PRONOSTIA and XJTU-SY datasets validate the effectiveness of the proposed method in improving the prediction accuracy of bearing health degradation indicators and enhancing the stability of multistep prediction in deep model.

A review of epilepsy detection and prediction methods based on EEG signal processing and deep learning

Epilepsy is a chronic neurological disorder that poses significant challenges to patients and their families. Effective detection and prediction of epilepsy can facilitate patient recovery, reduce family burden, and streamline healthcare processes. Therefore, it is essential to propose a deep learning method for efficient detection and prediction of epileptic electroencephalography (EEG) signals. This paper reviews several key aspects of epileptic EEG signal processing, focusing on epilepsy detection and prediction. It covers publicly available epileptic EEG datasets, preprocessing techniques, feature extraction methods, and deep learning-based networks used in these tasks. The literature is categorized based on patient independence, distinguishing between patient-independent and non-patient-independent studies. Additionally, the evaluation methods are classified into general classification indicators and specific epilepsy prediction criteria, with findings organized according to the prediction cycles reported in various studies. The review reveals several important insights. Despite the availability of public datasets, they often lack diversity in epilepsy types and are collected under controlled conditions that may not reflect real-world scenarios. As a result, signal preprocessing methods tend to be limited and may not fully represent practical conditions. Feature extraction and network designs frequently emphasize fusion mechanisms, with recent advances in Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) showing promising results, suggesting that new network models warrant further exploration. Studies using patient-independent data generally produce better results than those relying on non-patient-independent data. Metrics based on general classification methods typically perform better than those using specific epilepsy prediction criteria, though future research should focus on the latter for more accurate evaluation. Epilepsy prediction cycles are typically kept under 1 h, with most studies concentrating on intervals of 30 min or less.

DeepRSMA: a cross-fusion based deep learning method for RNA-small molecule binding affinity prediction.

RNA is implicated in numerous aberrant cellular functions and disease progressions, highlighting the crucial importance of RNA-targeted drugs. To accelerate the discovery of such drugs, it is essential to develop an effective computational method for predicting RNA-small molecule affinity (RSMA). Recently, deep learning based computational methods have been promising due to their powerful nonlinear modeling ability. However, the leveraging of advanced deep learning methods to mine the diverse information of RNAs, small molecules and their interaction still remains a great challenge. In this study, we present DeepRSMA, an innovative cross-attention-based deep learning method for RSMA prediction. To effectively capture fine-grained features from RNA and small molecules, we developed nucleotide-level and atomic-level feature extraction modules for RNA and small molecules, respectively. Additionally, we incorporated both sequence and graph views into these modules to capture features from multiple perspectives. Moreover, a Transformer-based cross-fusion module is introduced to learn the general patterns of interactions between RNAs and small molecules. To achieve effective RSMA prediction, we integrated the RNA and small molecule representations from the feature extraction and cross-fusion modules. Our results show that DeepRSMA outperforms baseline methods in multiple test settings. The interpretability analysis and the case study on spinal muscular atrophy (SMA) demonstrate that DeepRSMA has the potential to guide RNA-targeted drug design. The codes and data are publicly available at https://github.com/Hhhzj-7/DeepRSMA. Supplementary data are available at Bioinformatics online.

Data discretization impact on deep learning for missing value imputation of continuous data

In various fields of information examination, for example, AI, profoundly getting the hang of missing information is a typical issue. Missing qualities should be tended to since they can adversely affect the exactness and adequacy of prescient models. This research investigates how data discretization affects deep learning methods for filling the missing values in datasets with continuous features. They provide a unique method for imputing missing values using deep neural networks (DNNs) called extravagant expectation maximization-deep neural network (EEM-DNN). This approach discretizes continuous features into separate intervals initially. This is justified by treating the issue of missing value imputation as a classification work, with the missing values being considered a distinct class. A DNN, designed explicitly for imputation, is then trained using the discretized data. The expectation maximization concepts are incorporated into the network architecture, and as a result, the network iteratively improves its imputation predictions. They run comprehensive experiments on several datasets from different fields to gauge the efficacy of the suggested strategy. The effectiveness of EEM-DNN is compared to that of other imputation approaches, such as traditional imputation techniques and deep learning methods without data discretization. Our findings show that data discretization significantly enhances imputation accuracy. In terms of imputation accuracy and prediction performance on downstream tasks, the EEM-DNN method regularly performs better than alternative methods. It also examines if various discretization techniques affect the overall imputation process. They find that the trade-off between bias and variance in imputed data depends on the discretization method selected. This highlights the significance of choosing a suitable discretization approach depending on the unique properties of the dataset.

Two-stage deep reinforcement learning method for agile optical satellite scheduling problem

Two-stage deep reinforcement learning method for agile optical satellite scheduling problem

Advanced Modulation Formats for 400 Gbps Optical Networks and AI-Based Format Recognition

The integration of communication and sensing (ICAS) in optical networks is an inevitable trend in building intelligent, multi-scenario, application-converged communication systems. However, due to the impact of nonlinear effects, co-fiber transmission of sensing signals and communication signals can cause interference to the communication signals, leading to an increased bit error rate (BER). This paper proposes a noncoherent solution based on the alternate polarization chirped return-to-zero frequency shift keying (Apol-CRZ-FSK) modulation format to realize a 4 × 100 Gbps dense wavelength division multiplexing (DWDM) optical network. Simulation results show that compared to traditional modulation formats, such as chirped return-to-zero frequency shift keying (CRZ-FSK) and differential quadrature phase shift keying (DQPSK), this solution demonstrates superior resistance to nonlinear effects, enabling longer transmission distances and better transmission performance. Moreover, to meet the transmission requirements and signal sensing and recognition needs in future optical networks, this study employs the Inception-ResNet-v2 convolutional neural network model to identify three modulation formats. Compared with six deep learning methods including AlexNet, ResNet50, GoogleNet, SqueezeNet, Inception-v4, and Xception, it achieves the highest performance. This research provides a low-cost, low-complexity, and high-performance solution for signal transmission and signal recognition in high-speed optical networks designed for integrated communication and sensing.

A review on deep learning methods for heart sound signal analysis

IntroductionApplication of Deep Learning (DL) methods is being increasingly appreciated by researchers from the biomedical engineering domain in which heart sound analysis is an important topic of study. Diversity in methodology, results, and complexity causes uncertainties in obtaining a realistic picture of the methodological performance from the reported methods.MethodsThis survey paper provides the results of a broad retrospective study on the recent advances in heart sound analysis using DL methods. Representation of the results is performed according to both methodological and applicative taxonomies. The study method covers a wide span of related keywords using well-known search engines. Implementation of the observed methods along with the related results is pervasively represented and compared.Results and discussionIt is observed that convolutional neural networks and recurrent neural networks are the most commonly used ones for discriminating abnormal heart sounds and localization of heart sounds with 67.97% and 33.33% of the related papers, respectively. The convolutional neural network and the autoencoder network show a perfect accuracy of 100% in the case studies on the classification of abnormal from normal heart sounds. Nevertheless, this superiority against other methods with lower accuracy is not conclusive due to the inconsistency in evaluation.

Design of Travel Route Planning Model Based on Deep Learning

Tourism route planning is affected by the factors such as tourism destination and tourists’ preference, which leads to poor automatic matching of routes. A mathematical model of tourism route planning based on deep learning is proposed. Under the condition of total route constraints, according to the historical tourism preference information and prior information of tourists, a big data model for the distribution of personalized characteristics of tourists is established, spatial constraints and time constraints parameters are input, and the feature matching of geographic location information and tourist interest parameter information is carried out by using the deep learning method, and the statistical feature quantity of the parameters of the tourism route planning model is extracted. Under the control of the deep learning and geographic information data set, Carry out multi-constraint and multi-objective hierarchical analysis of tourists and tourism destinations, and realize the optimization design of tourism route planning algorithm. The simulation results show that the accuracy of this method is high and the deviation is small. It improves the satisfaction level of tourists and helps users complete a series of visual analysis tasks such as route mining, route planning and destination analysis.