Accuracy Of Classifier Research Articles

Exploring Binary Classification Models for Parkinson’s Disease Detection

Parkinson’s disease (PD) is a degenerative condition characterised by a combination of motor symptoms, such as tremors and rigidity, and non-motor symptoms, including cognitive impairment and anxiety. Although there is currently no known cure, there are available treatments that can effectively mitigate symptoms and enhance life expectancy. Our research primarily centred around vocal disorders, given that an estimated 90% of individuals diagnosed with Parkinson’s disease experience the onset of such disorders in the early stages of the disease. In this study, a dataset is collected from UCI repository consisting of 252 cases was utilised, with 188 cases classified as Parkinson’s disease and 64 cases classified as non-Parkinson’s disease. The results of our study indicate that the K-nearest neighbours (KNN) algorithm demonstrated a peak accuracy of 98.52% and an average accuracy of 97.33%, surpassing the performance of previously established models. Additionally, we introduced the implementation of Parzen Windows Estimation (a KDE technique) and computed classifier accuracies based on mean test set scores after multiple rounds of data reshuffling and model training. Our results enhance credibility compared to previous works in this field. Given the high misdiagnosis rate (26%) and the expense of imaging tests, deploying a precise and cost-effective algorithm for early PD detection is critical.

Analyzing the Use of Social Media Data to Understand Long-Term Crisis Management Challenges of COVID-19

In the past three years, social media has had a significant impact on our lives, including crisis management. The COVID-19 pandemic highlighted the importance of accurate information and exposed the spread of false information. This paper specifically examines the COVID-19 crisis and analyzes relevant literature to provide insights for national authorities and organizations. Utilizing social media data for crisis management poses challenges due to its unstructured nature. To overcome this, the paper proposes a comprehensive method that addresses all aspects of long-term crisis management. This method relies on labeled and structured information for accurate sentiment analysis and classification. An automated approach is presented to annotate and classify tweet texts, reducing manual labeling and improving classifier accuracy. The framework involves generating topics using Latent Dirichlet Allocation (LDA) and ranking them with a new algorithm for data annotation. The labeled text is transformed into feature representation using Bert embeddings, which can be utilized in deep learning models for categorizing textual data. The primary aim of this paper is to offer valuable insights and resources to researchers studying crisis management through social media literature, with a specific focus on high-accuracy sentiment analysis.

A Graph Convolution Model for Intelligent Datum Features Selection

The datum is a crucial component in tolerance specification, which is the foundation for the selections of geometric tolerances and tolerance principles. Currently, intelligent datum reasoning is largely based on logical rules that are mainly extracted from human experience, resulting in the high uncertainty and low efficiency. To tickle these issues, this study proposes a data selection model based on the GCN (Graph Convolutional Networks), In the devised model, the different geometric features of a workpiece are represented in a graph structure. The geometric, spatial, and assembly relationships, as well as positioning features are computed to obtain vectorized representations of the different geometric features, which serve as inputs to the constructed GCN model. Then, based on the GCN, a datum discriminant classifier has been developed on the training samples. To enhance the classifier accuracy, multiple GCN layers are employed for training, with the output of each GCN module added to a list. Ultimately, the outputs of all GCN modules are concatenated and subjected to classification prediction through fully connected layers. Datum specifications are established based on the classification of geometric features. The effectiveness and feasibility of this method are validated through case studies, e.g. rear floor crossbeams, with comparative results indicating a similarity rate of 85.19% with manually designed outcomes.

Key generation method from fingerprint image based on deep convolutional neural network model

Biometrics effect our live. Security applications employ biometrics. Biometric encryption is growing. Encryption requires biometric key creation. Long, random, and unexpected is the key. Information and communication security research emphasizes long, strong encryption keys. The proposed system uses fingerprint biometrics to generate a long, random biometric encryption key for symmetric encryption. Pre-processing removed noise from donor fingerprint images in the dataset. The program then trains an updateable Tuned VGG-16 convolutional neural network model and tests it on fingerprint images to learn fundamental fingerprint properties. The convolutional neural netwoprk CNN model retains the final weights for the second model to extract encryption key features. Transfer learning built a second convolutional neural network model to retrieve features without relearning. Keeping vector mean for processing. The last step generates an encryption key based on each person's vector of unique biometric features can be used for symmetric encryption algorithms to encrypt personal documents on the personal PC or personal cloud. Our CNN based method uses biometrics to recognize people and create safe and trustworthy encryption keys with over 99% accuracy in testing. Our 98%-accurate deep ANN classifier exceeds the support vector machine and random forest classifiers.

Implementation of Xception and EfficientNetB3 for COVID-19 Detection on Chest X-Ray Image via Transfer Learning

COVID-19 is a highly contagious infectious disease caused by the SARS-CoV-2 virus that can cause respiratory issues. The utilization of X-ray imaging has the potential to serve as an alternative means of detecting COVID-19 by offering insights into the condition of the lungs. Rapid and automated analysis of medical images and patterns can be achieved through deep learning techniques. In this study, we propose methods for the automatic classification of COVID-19 from Chest X-Ray images using CNN with transfer learning techniques, namely Xception and EfficientNetB3 architectures, as well as an ensemble of both architectures working in parallel. Additionally, we use Grad-CAM to visualize the regions of the X-ray image that are most important for the classification decision. The classification of COVID-19 is carried out for four types of classes: COVID-19, normal, bacterial pneumonia, and viral pneumonia. The proposed classifier models achieve an overall accuracy of 94.44% for the Xception classifier, 95.28% for the EfficientNetB3 classifier, and 94.44% for the parallel classifier. The accuracy value is higher than the other comparison classifiers accuracy values. Overall, the proposed classifiers can be recommended as tools to assist radiologists and clinical practitioners in diagnosing and following up with COVID-19 cases.

GHA-DenseNet prediction and diagnosis of malignancy in femoral bone tumors using magnetic resonance imaging

Background and objectiveDue to their aggressive nature and poor prognosis, malignant femoral bone tumors present considerable hurdles. Early treatment commencement is essential for enhancing vital and practical outcomes. In this investigation, deep learning algorithms will be used to analyze magnetic resonance imaging (MRI) data to identify bone tumors that are malignant. MethodologyThe study cohort included 44 patients, with ages ranging from 17 to 78 (22 women and 22 males). To categorize T1 and T2 weighted MRI data, this paper presents an improved DenseNet network model for the classification of bone tumor MRI, which is named GHA-DenseNet. Based on the original DenseNet model, the attention module is added to solve the problem that the deep convolutional model can reduce the loss of key features when capturing the location and content information of femoral bone tumor tissue due to the limitation of local receptive field. In addition, the sparse connection mode is used to prune the connection mode of the original model, so as to remove unnecessary and retain more useful fast connection mode, and alleviate the overfitting problem caused by small dataset size and image characteristics. In a clinical model designed to anticipate tumor malignancy, the utilization of T1 and T2 classifier output values, in combination with patient-specific clinical information, was a crucial component. ResultsThe T1 classifier's accuracy during the training phase was 92.88% whereas the T2 classifier's accuracy was 87.03%. Both classifiers demonstrated accuracy of 95.24% throughout the validation phase. During training and validation, the clinical model's accuracy was 82.17% and 81.51%, respectively. The clinical model's receiver operating characteristic (ROC) curve demonstrated its capacity to separate classes. ConclusionsThe proposed method does not require manual segmentation of MRI scans because it makes use of pretrained deep learning classifiers. These algorithms have the ability to predict tumor malignancy and shorten the diagnostic and therapeutic turnaround times. Although the procedure only needs a little amount of radiologists' involvement, more testing on a larger patient cohort is required to confirm its efficacy.

An Efficient Framework Based on Semi‐Supervised Learning for Transformer Fault Diagnosis

Transformer fault diagnosis is very important for the stable operation of transformers, but it is still difficult to recognize transformer faults. The main obstacle is attributed to the lack of labeled data for transformer faults. Due to the limited data of transformer faults, a great deal of unlabeled data in real life is not difficult to obtain and contains a lot of effective information. Semi‐supervised self‐training can significantly enhance the classification accuracy and performance of the base classifier by applying unlabeled data. However, one of the reasons for the degradation of semi‐supervised self‐training is the lack of initial labeled data to train the initial classifier. In this paper, we propose an improved semi‐supervised SMOTE (SS‐SMOTE), which uses the potential information contained in unlabeled data to generate synthetic samples, and then reasonably increases the labeled data set. Meanwhile, we combine SS‐SMOTE with self‐training into an algorithm framework. Finally, a series of experiments are carried out to evaluate the performance of the proposed algorithm, and compared with other existing state‐of‐the‐art semi‐supervised algorithms. The results show that the performance improvement of the classifier using the semi‐supervised improved SMOTE algorithm framework is effective, and has a good effect on solving the problem of small sample classification. © 2023 Institute of Electrical Engineer of Japan and Wiley Periodicals LLC.

Heterogeneity‐oriented ensemble learning for rail monitoring based on vehicle‐body vibration

AbstractIn this study, railway vehicle‐body vibration was applied to rail detection for convenient sensor deployment and cost‐effectiveness. However, the waveform is difficult to analyze due to damping and interference. Data‐driven methods can help concatenate multidimensional signals and complex rail‐surface irregularities but are impressionably uncertain. This study proposes a method in which a deep learning framework is coupled with heterogeneous factors at every link in its ensemble strategy. The instantiation, module foundation, and scenario description establish a concrete system for dealing with the dilemma that an insufficient database hinders feature extraction, causing unproductive capacity. The performance is quantitatively discussed by combining different levels of on‐site rail‐surface conditions, reaching prediction errors of 4.7% and 6.5% and classifier accuracies of 98.4% and 93.7% for irregularities and defect severities, respectively. This work describes a way to extend self‐learner applicability in industry and will facilitate new support for railway track management.

Differgram: A convex optimization-based method for extracting optimal frequency band for fault diagnosis of rotating machinery

The extraction of fault resonance bands from a full frequency band has always stood as a classical and effective strategy for fault diagnosis in rotating machinery. Among the existing techniques for capturing fault frequency bands, methods such as fast kurtogram and its variants have been prevalent. However, these methods encounter two main limitations. Firstly, they are subject to the constraints of index performance, potentially leading to the identification of incorrect frequency bands. Secondly, they do not fully harness the potential of healthy signals, thus hindering accurate fault frequency band localization. While approaches like SKRgram and accugram endeavor to leverage healthy signals to highlight differences between healthy and faulty signals, SKRgram falls short in thoroughly exploring the significance of healthy signals. Accugram relies on classifier accuracy to pinpoint fault frequency bands, rendering it less robust and less interpretable. To capitalize more effectively on healthy signals' potential, this study proposed a novel approach called differgram based on the convex optimization model. By inputting both healthy and faulty signals into the differgram, optimal fault frequency bands can be extracted. By solving the constructed convex optimization model, the optimal difference frequency bands between healthy and faulty signals are automatically identified. The method is substantiated through mathematical proofs and physical explanations, enhancing its interpretability. Through simulations and experimental validation, the proposed diffetergram demonstrates its efficiency in the extraction of fault frequency bands. In comparison to fast kurtogram, SKRgram, and accugram, the differgram showcases heightened robustness and noise immunity.

Automated machine learning and neural architecture optimization

Automated machine learning (AutoML) and neural architecture optimization (NAO) represent pivotal components in the landscape of machine learning and artificial intelligence. This paper extensively explores these domains, aiming to delineate their significance, methodologies, cutting-edge techniques, challenges, and emerging trends. AutoML streamlines and democratizes machine learning by automating intricate procedures, such as algorithm selection and hyperparameter tuning. Conversely, NAO automates the design of neural network architectures, a critical aspect for optimizing deep learning model performance. Both domains have made substantial advancements, significantly impacting research, industry practices, and societal applications. Through a series of experiments, classifier accuracy, NAO model selection based on hidden unit count, and learning curve analysis were investigated. The results underscored the efficacy of machine learning models, the substantial impact of architectural choices on test accuracy, and the significance of selecting an optimal number of training epochs for model convergence. These findings offer valuable insights into the potential and limitations of AutoML and NAO, emphasizing the transformative potential of automation and optimization within the machine learning field. Additionally, this study highlights the imperative for further research to explore synergies between AutoML and NAO, aiming to bridge the gap between model selection, architecture design, and hyperparameter tuning. Such endeavors hold promise in opening new frontiers in automated machine learning methodologies.

A model for identifying project sprint tasks based on their description

The subject of research in this article is the identification of project sprint tasks. The purpose of the article is to find approaches to reducing the risks of not fulfilling sprint tasks. The article solves the following tasks: analyzing research on the classification and visualization of project tasks, developing an algorithm that can automatically classify text descriptions of sprint tasks, collecting and preparing a training sample of text descriptions of sprint tasks for training and testing the classification model, applying natural language processing methods to improve classification and ensure the accuracy of the results, validating the model on real data to assess the efficiency and accuracy of classification, and analyzing the results. The following methods have been used: machine learning methods for classification, text vectorization methods, methods for classifying text descriptions, natural language processing methods, methods for semantic analysis of task description text, methods for processing expert opinions. The following results were obtained: a comprehensive approach to using machine learning algorithms, including the collection and processing of textual descriptions of tasks, for classification and involvement of expert opinions to improve the quality of task perception by the project team. Text expressions were classified based on the Bayesian classifier and neural classifiers. A visual representation of the data was implemented. Semantic analysis of the text of the description and title of the tasks was performed. Data markup was obtained to classify the quality of the wording, which was performed by a team of experts. To measure the reliability of the obtained expert assessments, we calculated Cohen's kappa coefficient for each pair of markers. According to the experimental results, the accuracy of the Bayesian classifier is 70%. For the classifier based on deep learning, a neural network for binary classification based on the transformer architecture was selected. The neural network was trained using the Python programming language and deep learning frameworks. The result is a classifier that gives an accuracy score of 83% on a test dataset, which is a good result for a small dataset and data with conflicting labels. Conclusions: The analysis of textual data confirms that the existing data in the tracking system is incomplete and contains abbreviations, conventions, and slang. The results show that the assessment of the quality of the wording is determined by the level of expert knowledge of the specifics and context of the project, while increasing the number of experts has almost no effect on the result. In further research, it is recommended to test the hypothesis that the effectiveness of the classifier depends on the specific project and the use of unsupervised learning methods for the task of identifying the quality of formulations.

Independent validation as a validation method for classification

The use of classifiers provides an alternative to conventional statistical methods. This involves using the accuracy with which data is correctly assigned to a given group by the classifier to apply tests to compare the performance of classifiers. The conventional validation methods for determining the accuracy of classifiers have the disadvantage that the distribution of correct classifications does not follow any known distribution, and therefore, the application of statistical tests is problematic. Independent validation circumvents this problem and allows the use of binomial tests to assess the performance of classifiers. However, independent validation accuracy is subject to bias for small training datasets. The present study shows that a hyperbolic function can be used to estimate the loss in classifier accuracy for independent validation. This function is used to develop three new methods to estimate the classifier accuracy for small training sets more precisely. These methods are compared to two existing methods in a simulation study. The results indicate overall small errors in the estimation of classifier accuracy and indicate that independent validation can be used with small samples. A least square estimation approach seems best suited to estimate the classifier accuracy.

Performance Augmentation of Base Classifiers Using Adaptive Boosting Framework for Medical Datasets

This paper investigates the performance enhancement of base classifiers within the AdaBoost framework applied to medical datasets. Adaptive boosting (AdaBoost), being an instance of boosting, combines other classifiers to enhance their performance. We conducted a comprehensive experiment to assess the efficacy of twelve base classifiers with the AdaBoost framework, namely, Bayes network, decision stump, ZeroR, decision tree, Naïve Bayes, J-48, voted perceptron, random forest, bagging, random tree, stacking, and AdaBoost itself. The experiments are carried out on five datasets from the medical domain based on various types of cancers, i.e., global cancer map (GCM), lymphoma-I, lymphoma-II, leukaemia, and embryonal tumours. The evaluation focuses on the accuracy, precision, and efficiency of the base classifiers in the AdaBoost framework. The results show that the performance of Naïve Bayes, Bayes network, and voted perceptron is highly improved compared to the rest of the base classifiers, attaining accuracies as high as 94.74%, 97.78%, and 97.78%, respectively. The results also show that in most cases, the base classifiers perform better with AdaBoost compared to their performance, i.e., for voted perceptron, the accuracy is improved up to 13.34%.For bagging, it is improved by up to 7%. This research aims to identify such base classifiers with optimal boosting capabilities within the AdaBoost framework for medical datasets. The significance of these results is that they provide insight into the performance of the base classifiers when used in the boosting framework to enhance the classification performance of classifiers in scenarios where individual classifiers do not perform up to the mark.

Improved fetal heartbeat detection using pitch shifting and psychoacoustics

ObjectiveFetal phonocardiography is a cost-effective and non-invasive method for monitoring of fetal health. However, the captured signal is often too noisy and unstable for reliable analysis. Detection of fetal heartbeats therefore requires a robust set of features to differentiate relevant sounds from background noise. We propose a novel set of features based on pitch shifting and psychoacoustics which lead to significant improvements of automatic fetal heartbeat detection. MethodsWe collected, annotated, cleaned, processed and segmented microphone recordings taken from pregnant women. In addition, we employed a well-known simulated dataset to validate the methodology further. A total of 212 features were extracted, including classical audio features, psychoacoustic features, and features based on Empirical Mode Decomposition. Methods from statistical analysis and machine learning were applied to demonstrate the relevance of specific features. Afterwards, we trained a set of machine learning algorithms with different subsets of features. ResultsPsychoacoustic features extracted from pitch-shifted signals are generally shown to be relevant and useful features for fetal heartbeat detection. When psychoacoustic features were added to the baseline features, classification accuracy of a robust classifier significantly increased from 69.24% to 76.16% for the real-world signals, and from 92.69% to 97.42% for one of the simulated datasets. ConclusionThe results demonstrate considerable gains in detection accuracy once pitch shifting and psychoacoustic modeling are applied to the input signal. SignificanceImprovements achieved from applying psychoacoustics on fetal phonocardiographic signals might be an important finding in the effort of making remote monitoring of fetal wellbeing inexpensive and accessible.

Analysis of an optical imaging system prototype for autonomously monitoring zooplankton in an aquaculture facility

Traditional approaches to biomonitoring in aquatic systems, such as sample collection, sorting, and identification, require significant time and effort, thereby limiting the spatiotemporal resolution of sample collection. Additionally, collection and preservation of samples for subsequent taxonomic identification and enumeration leads to mortality of organisms. Recent advances in technologies that utilize optical imaging and machine learning have provided new opportunities to expedite biomonitoring and lead to significant cost savings. These technologies can be advantageous to scientists or managers that conduct routine biomonitoring to inform operations, as in the case of aquaculture facilities. The Small Aquatic Organism optical imaging system (SAO) is a high-throughput optical imaging and classification prototype system that relies on computer vision and machine learning (Support Vector Machines, or SVMs) to autonomously identify and enumerate aquatic organisms. The SAO provides a more sustainable method of collecting large volumes of data and has the benefit of being used in situ. In this study, we tested the performance of the SAO in providing comparable results to manual zooplankton community monitoring in ten ponds at an aquaculture facility. We performed a side-by-side study comparing the sampling methods of plankton tow nets, where major zooplankton taxonomic classes were manually identified and enumerated, to sampling with the SAO. Vouchered samples were used to develop a training library for the SAO, where classes consisted of water boatman and zooplankton groups: cladocerans, copepod adults, copepod nauplii, and rotifers. SAO imagery was manually classified and compared with predicted results for validation. Accuracy for the SVM classifier of the SAO was 37.4 %. Convolutional Neural Networks (CNN) and Random Forest classifiers were also applied to SAO imagery and image features for comparison. The best CNN model and our Random Forest model had accuracies of 80.4 % and 46.6 % respectively. Challenges faced included the small size of copepod nauplii and rotifers and the limited resolution of the imaging camera, although there are tradeoffs between imaging resolution and the sample processing rate. Our comparison shows that advancement in both optical imaging and ML are needed in order for the SAO prototype to yield comparable results to manual community monitoring in an aquaculture facility.

Quasi-movements and attempted movements: a possible alternative to motor imagery in BCI-based neurorehabilitation

Motor imagery (MI) is a frequently used “mental trigger” for non-invasive brain-computer interfaces (BCI). Numerous studies have examined the effectiveness of MI-BCI for post-stroke rehabilitation. However, the results remain inconclusive. A potential obstacle to the effectiveness of this method could stem from an ongoing debate between the internal focus of mental activity (i.e., modeling of reality) inherent in MI and the perceived significance of sensory feedback from the actual physical environment in BCI-facilitated therapy. The requirement to allocate attentional resources to both internal actions and external consequences may contribute to the low accuracy of MI-BCI classifiers in most users. Moreover, internal focus of attention in MI may partially account for the consistent failures in combining MI-BCI with eye tracker-based interaction technologies, since external focus of attention is crucial for gaze control. A potentially effective replacement for motor imagery in BCIs is attempted movements (AMs). Studies have shown that BCIs are more successful in decoding AMs than MI (e.g., [1]). AMs involve attempted, but unrealized movements caused by paralysis or amputation. Despite their potential, AMs have received little attention, possibly because of modeling challenges with healthy participants and the widespread popularity of MI-BCIs. One approach to modeling them in healthy subjects is to use quasi-movements (QM), which are voluntary movements that are minimized by the subject to such an extent that they eventually become undetectable by objective measures [2]. However, QM has been studied even less than AM since its discovery by V.V. Nikulin and colleagues [2]. This may be due to the inadequate understanding of the difference between QM and MI. We recently proved that the sensorimotor rhythm’s event-related desynchronization in QM does not rely on the residual electromyogram (EMG), indicating that strict EMG control, which is often impossible, may not be necessary for QM contrary to prior views. As a result, QM can be embraced more frequently as an alternative to MI in BCIs [3]. Moreover, we substantiated and refined earlier findings [2] that QM possesses a striking similarity to actual movements [4]. Here, we present our initial findings on the asynchronous classification of QM. These findings may serve as a foundation for a real-time QM-BCI system. We used the EEG data recorded from 23 participants who synchronized their QM and MI with rhythmic sound triplets. A convolutional neural network called SimpleNet [5], with high interpretability, was trained on a subset of individual data separately for QM and IM, compared to a referential non-motor task. The network was applied offline and was unaware of sound timing, to another subset in 1.5-second windows with 0.1-second steps. QM/IM were detected when four consecutive positive windows occurred with a refractory period of three seconds. Due to the high variability in MI-BCI performance among untrained individuals, we only assessed classifier performance amongst participants exhibiting a TPR (true positive rate) greater than 0.5. We identified 7 such participants in QM and 5 in MI. QM showed better intention detection than MI, though not significantly, according to the Mann-Whitney test. The mean values ± standard deviation for TPR were 0.81±0.12 in QM and 0.77±0.12 in MI, while the false alarm rate (s–1) was 0.03±0.02 in QM and 0.04±0.03 in MI, and the response time (s) was 2.81±0.06 in QM and 2.86±0.10 in MI. Our initial findings on asynchronous BCI modeling are consistent with previous studies that have demonstrated superior QM classification in synchronous paradigms as compared to MI [2]. Notably, only minor hyperparameter optimization was performed for SimpleNet, leaving ample room for improvement in classification. Given the superior classification of AM over MI [1] and the promising preliminary findings presented here, the use of AM by end-users appears to be a viable option. Utilizing QM in studies that model AM could likely lead to further promotion and development of this technique. Additionally, the comparable nature of QM, AM, and overt movement implies their applicability in conveying intention via gaze-controlled interfaces. Although overt motor confirmation is suitable, MI-BCIs have shown inadequacies in this regard.

IBJA: An improved binary DJaya algorithm for feature selection

Feature Selection (FS) is a special preprocessing step in Machine Learning (ML) that reduces the number of unwanted features in datasets to increase the accuracy of ML classifiers. A popular binary variant of the continuous Jaya algorithm is the Discrete Jaya (DJaya) algorithm. It is commonly used for addressing optimization problems with binary design variables (also known as binary decision variables). Nevertheless, DJaya tends to prematurely converge to local optimal solutions, and its performance deteriorates as the complexity of the optimization problem grows. The DJaya algorithm is improved in this article by introducing the Improved Binary DJaya Algorithm (IBJA), which is specially designed to solve the FS problem. IBJA includes three techniques in DJaya. First, it incorporates the update equation of the Harris Hawks Optimization (HHO) algorithm into the optimization loop of DJaya to enhance DJaya’s searching process and exploration abilities. Second, it uses a new Dynamic Opposition-based Learning in the final steps of the optimization loop of DJaya to boost its searching and exploring capabilities. Third, it employs binary transfer functions to calculate binary solutions from the real-valued solutions generated by HHO and DOBL. Using 15 UCI datasets, IBJA’s performance was assessed and compared against four ML classifiers and ten efficient optimization algorithms. Besides, the Friedman statistical test was employed to investigate the reliability of the experimental findings. According to the overall experimental and statistical findings, IBJA scored the highest accuracy, best objective value, and fewest chosen features for each of the 15 UCI datasets.

SAMPLER: unsupervised representations for rapid analysis of whole slide tissue images

Deep learning has revolutionized digital pathology, allowing automatic analysis of hematoxylin and eosin (H&E) stained whole slide images (WSIs) for diverse tasks. WSIs are broken into smaller images called tiles, and a neural network encodes each tile. Many recent works use supervised attention-based models to aggregate tile-level features into a slide-level representation, which is then used for downstream analysis. Training supervised attention-based models is computationally intensive, architecture optimization of the attention module is non-trivial, and labeled data are not always available. Therefore, we developed an unsupervised and fast approach called SAMPLER to generate slide-level representations. Slide-level representations of SAMPLER are generated by encoding the cumulative distribution functions of multiscale tile-level features. To assess effectiveness of SAMPLER, slide-level representations of breast carcinoma (BRCA), non-small cell lung carcinoma (NSCLC), and renal cell carcinoma (RCC) WSIs of The Cancer Genome Atlas (TCGA) were used to train separate classifiers distinguishing tumor subtypes in FFPE and frozen WSIs. In addition, BRCA and NSCLC classifiers were externally validated on frozen WSIs. Moreover, SAMPLER's attention maps identify regions of interest, which were evaluated by a pathologist. To determine time efficiency of SAMPLER, we compared runtime of SAMPLER with two attention-based models. SAMPLER concepts were used to improve the design of a context-aware multi-head attention model (context-MHA). SAMPLER-based classifiers were comparable to state-of-the-art attention deep learning models to distinguish subtypes of BRCA (AUC=0.911±0.029), NSCLC (AUC=0.940±0.018), and RCC (AUC=0.987±0.006) on FFPE WSIs (internal test sets). However, training SAMLER-based classifiers was >100 times faster. SAMPLER models successfully distinguished tumor subtypes on both internal and external test sets of frozen WSIs. Histopathological review confirmed that SAMPLER-identified high attention tiles contained subtype-specific morphological features. The improved context-MHA distinguished subtypes of BRCA and RCC (BRCA-AUC=0.921±0.027, RCC-AUC=0.988±0.010) with increased accuracy on internal test FFPE WSIs. Our unsupervised statistical approach is fast and effective for analyzing WSIs, with greatly improved scalability over attention-based deep learning methods. The high accuracy of SAMPLER-based classifiers and interpretable attention maps suggest that SAMPLER successfully encodes the distinct morphologies within WSIs and will be applicable to general histology image analysis problems. This study was supported by the National Cancer Institute (Grant No. R01CA230031 and P30CA034196).

Efficient knowledge distillation for remote sensing image classification: a CNN-based approach

PurposeThe paper aims to tackle the classification of Remote Sensing Images (RSIs), which presents a significant challenge for computer algorithms due to the inherent characteristics of clustered ground objects and noisy backgrounds. Recent research typically leverages larger volume models to achieve advanced performance. However, the operating environments of remote sensing commonly cannot provide unconstrained computational and storage resources. It requires lightweight algorithms with exceptional generalization capabilities.Design/methodology/approachThis study introduces an efficient knowledge distillation (KD) method to build a lightweight yet precise convolutional neural network (CNN) classifier. This method also aims to substantially decrease the training time expenses commonly linked with traditional KD techniques. This approach entails extensive alterations to both the model training framework and the distillation process, each tailored to the unique characteristics of RSIs. In particular, this study establishes a robust ensemble teacher by independently training two CNN models using a customized, efficient training algorithm. Following this, this study modifies a KD loss function to mitigate the suppression of non-target category predictions, which are essential for capturing the inter- and intra-similarity of RSIs.FindingsThis study validated the student model, termed KD-enhanced network (KDE-Net), obtained through the KD process on three benchmark RSI data sets. The KDE-Net surpasses 42 other state-of-the-art methods in the literature published from 2020 to 2023. Compared to the top-ranked method’s performance on the challenging NWPU45 data set, KDE-Net demonstrated a noticeable 0.4% increase in overall accuracy with a significant 88% reduction in parameters. Meanwhile, this study’s reformed KD framework significantly enhances the knowledge transfer speed by at least three times.Originality/valueThis study illustrates that the logit-based KD technique can effectively develop lightweight CNN classifiers for RSI classification without substantial sacrifices in computation and storage costs. Compared to neural architecture search or other methods aiming to provide lightweight solutions, this study’s KDE-Net, based on the inherent characteristics of RSIs, is currently more efficient in constructing accurate yet lightweight classifiers for RSI classification.

Applied machine learning for stroke differentiation by electrical impedance tomography with realistic numerical models

Electrical impedance tomography (EIT) may have potential to overcome existing limitations in stroke differentiation, enabling low-cost, rapid, and mobile data collection. Combining bioimpedance measurement technologies such as EIT with machine learning classifiers to support decision-making can avoid commonly faced reconstruction challenges due to the nonlinear and ill-posed nature of EIT imaging. Therefore, in this work, we advance this field through a study integrating realistic head models with clinically relevant test scenarios, and a robust architecture consisting of nested cross-validation and principal component analysis. Specifically, realistic head models are designed which incorporate the highly conductive layers of cerebrospinal fluid in the subarachnoid space and ventricles. In total, 135 unique models are created to represent a large patient population, with normal, haemorrhagic, and ischemic brains. Simulated EIT voltage data generated from these models are used to assess the classification performance of support vector machines. Parameters explored include driving frequency, signal-to-noise ratio, kernel function, and composition of binary classes. Classifier accuracy at 60 dB signal-to-noise ratio, reported as mean and standard deviation, are (79.92% ± 10.82%) for lesion differentiation, (74.78% ± 3.79%) for lesion detection, (77.49% ± 15.90%) for bleed detection, and (60.31% ± 3.98%) for ischemia detection (after ruling out bleed). The results for each method were obtained with statistics from 3 independent runs with 17,280 observations, polynomial kernel functions, and feature reduction of 76% by PCA (from 208 to 50 features). While results of this study show promise for stroke differentiation using EIT data, our findings indicate that the achievable accuracy is highly dependent on the classification scenario and application-specific classifiers may be necessary to achieve acceptable accuracy.