Significant Increase In Classification Accuracy Research Articles

Automated classification of MRI images for mild cognitive impairment and Alzheimer’s disease detection using log gabor filters based deep meta-heuristic classification

In older people, mild cognitive impairment (MCI) is a precursor to more severe forms of dementia like AD (AD). In diagnosing patients with primary AD and amnestic MCI, modern neuroimaging techniques, especially MRI, play a key role. To efficiently categorize MRI images as normal or abnormal, the research presents a machine learning-based automatic labelling system, with a focus on boosting performance via texture feature analysis. To this end, the research implements a preprocessing phase employing Log Gabor filters, which are particularly well-suited for spatial frequency analysis. In addition, the research uses Gray Wolf Optimization (GWO) to acquire useful information from the images. For classification tasks using the MRI images, the research also make use of DenseNets, a form of deep neural network. The proposed method leverages Log Gabor filters for preprocessing, Gray Wolf Optimization (GWO) for feature extraction, and DenseNets for classification, resulting in a robust approach for categorizing MRI images as normal or abnormal. When compared to earlier trials performed without optimization, the proposed systematic technique shows a significant increase in classification accuracy of 15%. For neuroimaging applications, our research emphasizes the use of Log Gabor filters for preprocessing, GWO for feature extraction, and DenseNets for classification, which can help with the early detection and diagnosis of MCI and AD.

Benchmarking common uncertainty estimation methods with histopathological images under domain shift and label noise.

In the past years, deep learning has seen an increase in usage in the domain of histopathological applications. However, while these approaches have shown great potential, in high-risk environments deep learning models need to be able to judge their uncertainty and be able to reject inputs when there is a significant chance of misclassification. In this work, we conduct a rigorous evaluation of the most commonly used uncertainty and robustness methods for the classification of Whole Slide Images, with a focus on the task of selective classification, where the model should reject the classification in situations in which it is uncertain. We conduct our experiments on tile-level under the aspects of domain shift and label noise, as well as on slide-level. In our experiments, we compare Deep Ensembles, Monte-Carlo Dropout, Stochastic Variational Inference, Test-Time Data Augmentation as well as ensembles of the latter approaches. We observe that ensembles of methods generally lead to better uncertainty estimates as well as an increased robustness towards domain shifts and label noise, while contrary to results from classical computer vision benchmarks no systematic gain of the other methods can be shown. Across methods, a rejection of the most uncertain samples reliably leads to a significant increase in classification accuracy on both in-distribution as well as out-of-distribution data. Furthermore, we conduct experiments comparing these methods under varying conditions of label noise. Lastly, we publish our code framework to facilitate further research on uncertainty estimation on histopathological data.

A Hybrid GCN and Filter-Based Framework for Channel and Feature Selection: An fNIRS-BCI Study

In this study, a channel and feature selection methodology is devised for brain-computer interface (BCI) applications using functional near-infrared spectroscopy (fNIRS). A graph convolutional network (GCN) is employed to select the appropriate and correlated fNIRS channels. Furthermore, in the feature extraction phase, the performance of two filter-based feature selection algorithms, (i) the minimum redundancy maximum relevance (mRMR) and (ii) ReliefF, is investigated. The five most commonly used temporal statistical features (i.e., mean, slope, maximum, skewness, and kurtosis) are used, whereas the conventional support vector machine (SVM) is utilized as a classifier for training and testing. The proposed methodology is validated using an available online dataset of motor imagery (left- and right-hand), mental arithmetic, and baseline tasks. First, the efficacy of the proposed methodology is shown for two-class BCI applications (i.e., left- vs. right-hand motor imagery and mental arithmetic vs. baseline). Second, the proposed framework is applied to four-class BCI applications (i.e., left- vs. right-hand motor imagery vs. mental arithmetic vs. baseline). The results show that the number of appropriate channels and features was significantly reduced, resulting in a significant increase in classification accuracy for both two-class and four-class BCI applications, respectively. Furthermore, both mRMR (i.e., 87.8% for motor imagery, 87.1% for mental arithmetic, and 78.7% for four-class) and ReliefF (i.e., 90.7% for motor imagery, 93.7% for mental arithmetic, and 81.6% for four-class) yielded high average classification accuracy p < 0.05 . However, the results of the ReliefF algorithm are more stable and significant.

Batch normalization embeddings for deep domain generalization

Domain generalization aims at training machine learning models to perform robustly across different and unseen domains. Several methods train models from multiple datasets to extract domain-invariant features, hoping to generalize to unseen domains. Instead, first we explicitly train domain-dependent representations leveraging ad-hoc batch normalization layers to collect independent domain’s statistics. Then, we propose to use these statistics to map domains in a shared latent space, where membership to a domain is measured by means of a distance function. At test time, we project samples from an unknown domain into the same space and infer properties of their domain as a linear combination of the known ones. We apply the same mapping strategy at training and test time, learning both a latent representation and a powerful but lightweight ensemble model. We show a significant increase in classification accuracy over current state-of-the-art techniques on popular domain generalization benchmarks: PACS, Office-31 and Office-Caltech.

Classification of Partial Discharges Originating From Multilevel PWM Using Machine Learning

A partial discharge (PD) is a small, localized breakdown which often appears in high-voltage insulation systems. PDs can be initiated within material defects such as voids or cracks when exposed to sufficiently high electric field levels. In this research work, we study and propose a machine learning (ML) approach to detect and classify PDs originating from multilevel pulsewidth modulation (PWM) waveforms as utilized in power electronic devices, such as inverters and variable frequency drives. Due to the increased use of power conversion units, new types of higher magnitude PDs have been observed to increase insulation degradation. By creating a cavity placed at different locations within the test object and exposing it to PWM voltage waveforms with two different rise times, a total of 345 660 PD events are recorded and organized into ten different classes. The maximum PD amplitude, duration, the time distance between consecutive PDs, and the area under the PD are used as features for PD classification. A unique way of concatenating a sequence of the extracted features to capture the temporal dependence of consecutive PDs is also presented. It is observed that when creating a sequence of information from consecutive observed PDs, a significant increase in classification accuracy can be obtained. Trained classifiers based on ensemble bagged decision trees (DTs) and long short-term memory (LSTM) architecture resulted in 95.3% and 98.5% average classification accuracy on test data respectively.

EEG Classification of Imaginary Left and Right Leg Movements based on Multivariate Empirical Mode Decomposition and Non-Linear Support Vector Machine

Motor-imagery brain-computer interfaces, as rehabilitation tools for motor-disabled individuals, could inherently enrich neuroplasticity and subsequently restore mobility. However, this endeavour's significant challenge is classifying left and right leg motor imagery tasks from non-stationary EEG signals. A subject-independent feature extraction method is essential in a BCI system, and this work involves developing a subject-independent algorithm to classify left/right leg motion intention. The Multivariate Empirical Mode Decomposition was used to decompose EEG during left and right foot movements during imagery tasks. We validated our proposed algorithm using open-access motor imagery data to detect the user's mental intention from EEG. Five subjects of various performance categories with almost 150 trials for each left/right leg MI of hand/leg/tongue, HaLT Paradigm, utilizing C3, C4, and Cz channels were examined to generalize this study to all subjects. A set of statistical features were extracted from the intrinsic mode functions, and the most relevant features were selected for classification using Sequential Floating Feature Selection. Different classifiers were trained using extracted features, and their performances' were evaluated. The findings suggest that the non-linear support vector machine is the best classification model, resulting in the mean classification sensitivity, specificity, precision, negative predictive value, F-measure, 98.15%, 90.74%, 91.97%, 98.33%, 94.72%, 94.44%, respectively. The proposed subject-independent signal processing method significantly improved the offline calibration mode by eliminating the frequency selection step, making it the common-used method for different types of MI-based BCI participants. Offline evaluations suggest that it can lead to significant increases in classification accuracy in comparison to current approaches.

The Integration of Multi-Source Remotely Sensed Data with Hierarchically Based Classification Approaches in Support of the Classification of Wetlands

Methodologies were developed to classify wetlands (Open Bog, Treed Bog, Open Fen, Treed Fen, and Swamps) from remotely sensed data using advanced classification algorithms through two hierarchical approaches. The data utilized included multispectral optical and thermal data (Landsat-5, and Landsat-8), radar imagery (Sentinel-1), and a digital elevation model. Goals were to determine the best way to combine imagery to classify wetlands through hierarchically based classification approaches to produce more accurate and efficient maps compared to standard classification. Algorithms used were Random Forest (RF), and Naïve Bayes. A hierarchically based RF classification methodology produced the most accurate classification result (91.94%). The hierarchically based approaches also improved classification accuracies for low-quality data, as defined through feature analysis, when compared to a nonhierarchical classifier. The hierarchical approaches also produced a significant increase in classification accuracy for the Naïve Bayes classifier versus the standard approach (∼12% increase) while not significantly increasing computation time – comparable in accuracy to the RF tests for around 20% the computational effort. Preselection of spectral bands, polarizations and other input parameters (Normalized Difference Vegetation Index, Normalized Difference Water Index, albedo, slope, etc.) using log-normal or RF variable importance analysis was very effective at identifying low-quality features and features which were of higher quality.

Wine Ontology Influence in a Recommendation System

Wine is the second most popular alcoholic drink in the world behind beer. With the rise of e-commerce, recommendation systems have become a very important factor in the success of business. Recommendation systems analyze metadata to predict if, for example, a user will recommend a product. The metadata consist mostly of former reviews or web traffic from the same user. For this reason, we investigate what would happen if the information analyzed by a recommendation system was insufficient. In this paper, we explore the effects of a new wine ontology in a recommendation system. We created our own wine ontology and then made two sets of tests for each dataset. In both sets of tests, we applied four machine learning clustering algorithms that had the objective of predicting if a user recommends a wine product. The only difference between each set of tests is the attributes contained in the dataset. In the first set of tests, the datasets were influenced by the ontology, and in the second set, the only information about a wine product is its name. We compared the two test sets’ results and observed that there was a significant increase in classification accuracy when using a dataset with the proposed ontology. We demonstrate the general applicability of the methodology to other cases, applying our proposal to an Amazon product review dataset.

Generative adversarial network-based electromagnetic signal classification: A semi-supervised learning framework

Generative adversarial network (GAN) has achieved great success in many fields such as computer vision, speech processing, and natural language processing, because of its powerful capabilities for generating realistic samples. In this paper, we introduce GAN into the field of electromagnetic signal classification (ESC). ESC plays an important role in both military and civilian domains. However, in many specific scenarios, we can't obtain enough labeled data, which cause failure of deep learning methods because they are easy to fall into over-fitting. Fortunately, semi-supervised learning (SSL) can leverage the large amount of unlabeled data to enhance the classification performance of classifiers, especially in scenarios with limited amount of labeled data. We present an SSL framework by incorporating GAN, which can directly process the raw in-phase and quadrature (IQ) signal data. According to the characteristics of the electromagnetic signal, we propose a weighted loss function, leading to an effective classifier to realize the end-to-end classification of the electromagnetic signal. We validate the proposed method on both public RML2016.04c dataset and real-world Aircraft Communications Addressing and Reporting System (ACARS) signal dataset. Extensive experimental results show that the proposed framework obtains a significant increase in classification accuracy compared with the state-of-the-art studies.

Assessing impact of channel selection on decoding of motor and cognitive imagery from MEG data

Objective. Magnetoencephalography (MEG) based brain–computer interface (BCI) involves a large number of sensors allowing better spatiotemporal resolution for assessing brain activity patterns. There have been many efforts to develop BCI using MEG with high accuracy, though an increase in the number of channels (NoC) means an increase in computational complexity. However, not all sensors necessarily contribute significantly to an increase in classification accuracy (CA) and specifically in the case of MEG-based BCI no channel selection methodology has been performed. Therefore, this study investigates the effect of channel selection on the performance of MEG-based BCI. Approach. MEG data were recorded for two sessions from 15 healthy participants performing motor imagery, cognitive imagery and a mixed imagery task pair using a unique paradigm. Performance of four state-of-the-art channel selection methods (i.e. Class-Correlation, ReliefF, Random Forest, and Infinite Latent Feature Selection were applied across six binary tasks in three different frequency bands) were evaluated in this study on two state-of-the-art features, i.e. bandpower and common spatial pattern (CSP). Main results. All four methods provided a statistically significant increase in CA compared to a baseline method using all gradiometer sensors, i.e. 204 channels with band-power features from alpha (8–12 Hz), beta (13–30 Hz), or broadband (α + β) (8–30 Hz). It is also observed that the alpha frequency band performed better than the beta and broadband frequency bands. The performance of the beta band gave the lowest CA compared with the other two bands. Channel selection improved accuracy irrespective of feature types. Moreover, all the methods reduced the NoC significantly, from 204 to a range of 1–25, using bandpower as a feature and from 15 to 105 for CSP. The optimal channel number also varied not only in each session but also for each participant. Reducing the NoC will help to decrease the computational cost and maintain numerical stability in cases of low trial numbers. Significance. The study showed significant improvement in performance of MEG-BCI with channel selection irrespective of feature type and hence can be successfully applied for BCI applications.

Predicting visual perceivability of scene objects through spatio-temporal modeling of retinal receptive fields

Retinal processing of a visual scene is an essential step of human visual perception. Although foveal vision is linked to the visual attention, perception is by not means limited to this region. Rather, the retinal field of view ranges from 60° nasal to 107° temporal, and from 70° superior to 80° inferior. Whether a scene object is visually perceived depends on both its visual appearance as well as its retinal location. We present a framework to evaluate the visual stimulus of a scene object with regard to different types of retinal receptive fields. Driven by gaze location provided by an eye tracker, the estimated retinal response considers the visual appearance of the object, its eccentricity in the users field of view, and the capabilities at the local retinal region. A desktop experiment shows that, in additional to foveal processing, the estimated retinal response leads to a significant increase in classification accuracy in terms of whether an object is reported as perceived by the user.

Detection of marginal ice zone in Synthetic Aperture Radar imagery using curvelet-based features: a case study on the Canadian East Coast

Monitoring the marginal ice zone (MIZ) is becoming increasingly important due to recent evidence that the width of the MIZ is changing with climate. A method to automatically detect the MIZ in synthetic aperture radar (SAR) imagery is proposed. The method utilizes the curve-like features of MIZ in SAR images. A multiscale strategy, the curvelet transform, is chosen to extract features from the SAR images. The statistical and co-occurrence features of curvelet coefficients at an appropriate scale are used to identify the MIZ from open water and consolidated ice. Experimental results show a significant increase in classification accuracy (89.7%) compared with the most commonly used MIZ definition from passive microwave sea ice concentration (74%), especially in the diffuse MIZ.

Analysis of predictor variables for mosquito species identification from dual-wavelength polarization-sensitive lidar measurements.

Mosquito-borne diseases are a major challenge for Human health as they affect nearly 700 million people every year. Monitoring insects is generally done through trapping methods that are tedious to set up, costly and present scientific biases. Entomological lidars are a potential solution to remotely count and identify mosquito species and gender in real-time. In this contribution, a dual-wavelength polarization sensitive lidar is used in laboratory conditions to retrieve the wingbeat frequency as well as optical properties of flying mosquitoes transiting through the laser beam. From the lidar signals, predictive variables are retrieved and used in a Bayesian classification. This paper focuses on determining the relative importance of the predictive variables used in the classification. Results show a strong dominance of the wingbeat frequency, the impact of predictive variables based on depolarization and backscattering ratios are discussed, showing a significant increase in classification accuracy.

Assessing the Impact of Spectral Resolution on Classification of Lowland Native Grassland Communities Based on Field Spectroscopy in Tasmania, Australia

This paper presents a case study for the analysis of endangered lowland native grassland communities in the Tasmanian Midlands region using field spectroscopy and spectral convolution techniques. The aim of the study was to determine whether there was significant improvement in classification accuracy for lowland native grasslands and other vegetation communities based on hyperspectral resolution datasets over multispectral equivalents. A spectral dataset was collected using an ASD Handheld-2 spectroradiometer at Tunbridge Township Lagoon. The study then employed a k-fold cross-validation approach for repeated classification of a full hyperspectral dataset, a reduced hyperspectral dataset, and two convoluted multispectral datasets. Classification was performed on each of the four datasets a total of 30 times, based on two different class configurations. The classes analysed were Themeda triandra grassland, Danthonia/Poa grassland, Wilsonia rotundifolia/Selliera radicans, saltpan, and a simplified C3 vegetation class. The results of the classifications were then tested for statistically significant differences using ANOVA and Tukey’s post-hoc comparisons. The results of the study indicated that hyperspectral resolution provides small but statistically significant increases in classification accuracy for Themeda and Danthonia grasslands. For other classes, differences in classification accuracy for all datasets were not statistically significant. The results obtained here indicate that there is some potential for enhanced detection of major lowland native grassland community types using hyperspectral resolution datasets, and that future analysis should prioritise good performance in these classes over others. This study presents a method for identification of optimal spectral resolution across multiple datasets, and constitutes an important case study for lowland native grassland mapping in Tasmania.

Ensemble-based canonical correlation forest (CCF) for land use and land cover classification using sentinel-2 and Landsat OLI imagery

ABSTRACTClassifying spectrally similar crop types in fragmented landscapes is a difficult task due to the low spectral and spatial resolution of satellite imagery. The objective of this study is twofold: (I) to evaluate the performance of a recent ensemble methodology, namely canonical correlation forest (CCF), and (ii) to investigate the potential of recently launched Sentinel-2 (S-2) image for crop classification. The algorithm is based on building a forest-type ensemble model using multiple canonical correlation trees. Its performance was compared to widely-used random forest (RF) and rotation forest (RotFor) for the classification of 4-band and 10-band S-2 and Landsat-8 (L-8) OLI datasets. In addition to overall accuracy estimations, results of the methods were assessed using recently introduced map-level and category-level disagreement measures. Results showed that the CCF and RotFor algorithms produced statistically similar results for the S-2 datasets but statistically different results for the OLI image whilst outperforming the RF algorithm for all cases. The CCF algorithm was found less sensitive to the ensemble size (i.e. number of trees) compared to RF algorithm. Also, for the S-2 data inclusion of six spectral bands at 20 m resolution resulted in a significant increase in classification accuracy (up to 6%).

Prediction of Primary Tumors in Cancers of Unknown Primary.

A cancer of unknown primary (CUP) is a metastatic cancer for which standard diagnostic tests fail to identify the location of the primary tumor. CUPs account for 3–5% of cancer cases. Using molecular data to determine the location of the primary tumor in such cases can help doctors make the right treatment choice and thus improve the clinical outcome. In this paper, we present a new method for predicting the location of the primary tumor using gene expression data: locating cancers of unknown primary (LoCUP). The method models the data as a mixture of normal and tumor cells and thus allows correct classification even in impure samples, where the tumor biopsy is contaminated by a large fraction of normal cells. We find that our method provides a significant increase in classification accuracy (95.8% over 90.8%) on simulated low-purity metastatic samples and shows potential on a small dataset of real metastasis samples with known origin.

Combining Airborne Laser Scanning and Aerial Imagery Enhances Echo Classification for Invasive Conifer Detection

The spread of exotic conifers from commercial plantation forests has significant economic and ecological implications. Accurate methods for invasive conifer detection are required to enable monitoring and guide control. In this research, we combined spectral information from aerial imagery with data from airborne laser scanning (ALS) to develop methods to identify invasive conifers using remotely-sensed data. We examined the effect of ALS pulse density and the height threshold of the training dataset on classification accuracy. The results showed that adding spectral values to the ALS metrics/variables in the training dataset led to significant increases in classification accuracy. The most accurate models (kappa range of 0.773–0.837) had either four or five explanatory variables, including ALS elevation, the near-infrared band and different combinations of ALS intensity and red and green bands. The best models were found to be relatively invariant to changes in pulse density (1–21 pls/m2) or the height threshold (0–2 m) used for the inclusion of data in the training dataset. This research has extended and improved the methods for scattered single tree detection and offered valuable insight into campaign settings for the monitoring of invasive conifers (tree weeds) using remote sensing approaches.

Discovering Differences in Gender-related Skeletal Muscle Aging through the Majority Voting-Based Identification of Differently Expressed Genes

Understanding gene function (GF) is still a significant challenge in system biology. Previously, several machine learning and computational techniques have been used to understand GF. However, these previous attempts have not produced a comprehensive interpretation of the relationship between genes and differences in both age and gender. Although there are several thousands of genes, very few differentially expressed genes play an active role in understanding the age and gender differences. The core aim of this study is to uncover new biomarkers that can contribute towards distinguishing between male and female according to the gene expression levels of skeletal muscle (SM) tissues. In our proposed multi-filter system (MFS), genes are first sorted using three different ranking techniques (t-test, Wilcoxon and Receiver Operating Characteristic (ROC)). Later, important genes are acquired using majority voting based on the principle that combining multiple models can improve the generalization of the system. Experiments were conducted on Micro Array gene expression dataset and results have indicated a significant increase in classification accuracy when compared with existing system.

Improved decoding of attentional selection in a cocktail party environment with EEG via automatic selection of relevant independent components.

Recently it has been shown to be possible to ascertain which speaker a subject is attending to in a cocktail party environment from single-trial (~60s) electroencephalography (EEG) data. The attentional selection of most of subjects could be decoded with a very high accuracy (>90%). However, the performance of many subjects fell below what would be required for a potential brain computer interface (BCI). One potential reason for this is that activity related to the stimuli may have a lower signal-to-noise ratio on the scalp for some subjects than others. Independent component analysis (ICA) is a commonly used method for denoising EEG data. However, its effective use often requires the subjective choice of the experimenter to determine which independent components (ICs) to retain and which to reject. Algorithms do exist to automatically determine the reliability of ICs, however they provide no information as to their relevance for the task at hand. Here we introduce a novel method for automatically selecting ICs which are relevant for decoding attentional selection. In doing so, we show a significant increase in classification accuracy at all test data durations from 60s to 10s. These findings have implications for the future development of naturalistic and user-friendly BCIs, as well as for smart hearing aids.

Satellite-based stand-wise forest cover type mapping using a spatially adaptive classification approach

Due to high variation in forest communities, forest structure and the fragmentation of the forested area in Central Europe, satellite-based forest inventory methods have to meet particularly high-quality requirements. This study presents an innovative method to combine official forest inventory information at stand level with multidate satellite imagery using a spatially adaptive classification approach for producing wall-to-wall forest cover maps of important tree species and management classes across multiple ownership regions in a heterogeneous low mountain range in Germany. The classification approach was applied to a 5,200-km2 area (about 2,080 km2 of forest land, mostly mixed forests) located in the Eifel mountain range in Central Europe. In comparison with conventional classifiers, our results demonstrate a significant increase in classification accuracy in the order of 12%. The method was tested with ASTER images but holds the potential to be used for regular state forest inventories based on standard and novel earth observation data supplied for instance from the SPOT-5 and RapidEye sensors.