High Test Accuracy Research Articles

Predicting the Physical Properties of Dark Matter Subhalos from Baryonic Parameters Using Machine Learning

Dark matter subhalos play an important role in galaxy formation and evolution. However, accurate prediction of dark matter properties remains a challenge of modern-day astronomy. In recent times, machine learning (ML) tools have shown promising results in solving numerous astrophysical problems. In this paper, we use data from the EAGLE simulations to determine the total mass and the half-mass radius of dark matter subhalos using structural properties of gas, star, black hole, and photometric features using gradient boosted decision trees (GBDT) and dense neural network. GBDT does not require data preprocessing, and results in better performance compared to the neural network. According to GBDT, the most important feature for subhalo radius and mass estimation is gas radius and black hole mass respectively. The all-features combined approach results in the highest test accuracy — Pearson’s correlation coefficient = 0.947 and 0.981, coefficient of determination = 0.898 and 0.962, normalized median absolute deviation = 0.111 and 0.114 for radius and mass respectively. We evaluate our model for masses and redshifts beyond its training range and find that GBDT demonstrates significantly better extrapolation capabilities than the neural network. We also test our model on simulations with different resolutions, and find that the discrepancies lie within 10% if the resolution is changed. This novel study incorporates the structural parameters of gas and black hole to determine the dark matter properties using a ML-based approach. The promising results of this study prove that ML tools can improve our current understanding of dark matter, and answer some of the basic cosmological questions.

Prediction of Potential Urban Hot Spots in Urban Environments Using Convolutional Neural Networks and Remote Sensing Techniques

Abstract. With the increase in built-up areas and rising urban populations, Land Surface Temperature (LST) has significantly increased, leading to the proliferation of Urban Hot Spots (UHS) in urban environments. To mitigate UHS proactively, researchers have conducted studies using various models to predict LST. However, current predictions are primarily based on data samples from isolated stations, making them unfeasible for continuous LST prediction on larger scales, such as regional levels. Therefore, this research aims to use Singapore as a case study to predict UHS on a regional scale using machine learning based on essential variables. Specifically, this research proposes training a Convolutional Neural Network (CNN) model using identified independent variables, including elevation, Normalized Difference Built-up Index (NDBI), Normalized Difference Moisture Index (NDMI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), population, and Land Use and Land Cover (LULC), along with the target variable, UHS. After training, the model achieves high test accuracy and is fed projected data, subsequently producing projected UHS locations. The findings indicate that new UHS primarily appear in southwestern areas, Marina Bay, and the northwest regions of the country. This research predicts that 2.24 percent of the site could be classified as UHS by 2025, compared to the current percentage of 0.95 percent. Based on these projections, the research proposes preventative measures to proactively mitigate UHS. This research fills the gap by constructing a prediction model that can predict UHS locations on a regional scale.

A-327 Statistical Technique-based Image Analysis for Digital Reading of Immunochemistry Cassette

Abstract Background Covid-19 transformed healthcare and how we think about clinical diagnostics. People prefer at home testing options that enable them to focus on staying healthy at their own comfort and privacy. Fast and accurate lateral flow assays or rapid diagnostic test (RDT) meets that need, however ensuring high test specificity and accuracy are of utmost importance. Visual readout of RDTs is subjective and time consuming. With the current technological advancement, medical devices are smarter and can be used for automated readout. The core of such devices are the algorithms that process the information despite several limitations and unknown variations affecting input signals. Many such algorithms are developed and are used for similar applications. Here, a statistical model for any optical device that is designed to automatically read an RDT with one or more test lines is described. Methods The described LFA algorithm is a innovative way to analyze images acquired with Complementary Metal-Oxide-Semiconductor (CMOS) based optical system with low Signal-to-noise ratio (SNR), giving a semi quantitative result. This also takes care of variations in illumination sources with relatively reduced computation, by converting the images to a 1 Dimensional (1D) signal. Algorithm workflow: Image pre-processing: Affine transformations such as rotation, translation, etc. (depends on the inherent setup) are applied on the acquired images as a pre-processing step. Pattern matching techniques using Gaussian templates are applied to determine the window region of the LFA and precisely identify the test and control line regions. Image formatting: The sample region of the cassette is converted from RGB color space to the weighted gray (Gw) for digital interpretation of the primary biomarker present. Characterization of the Gw channel enables test peak detection with enhanced resolution. This helps in early call out of results and improved accuracy, thereby reducing chances of erroneous results. A second order Butterworth low pass filter is applied to reduce system noise and enhance SNR. The images are converted to 1D signal by averaging for mathematical read out. Mathematical interpretation: Maximum peak height (max) and minimum peak depth (min) of Gw channel values in the test line region (t-peak) are extracted from the 1D signal. The mathematical representation of the test line is calculated as: Test Peak Signal = {max(t-peak) - min(t-peak)} / Alpha; Normalization Factor, Alpha = (sum of all Gw channel values for any clear non-reactive part of the test device) / (number of Gw channel values) Results TPS values for various biomarkers from statistically significant replicates were computed at Low, Medium, and High levels. Two sample T-tests were performed, and respective p values are distinctly different as p-value<0.05. For a concentration level of Low to medium, p-value is 0.044 and for Medium - High p-value is 0.004. Conclusions This algorithm can run on any off-the-shelf available RDT coupled with a CMOS based optical system. The proposed algorithm can differentiate between low, medium, and high concentrations with high confidence. High throughput semiquantitative digital readouts are robust, accurate and are preferred for diagnostic test result interpretation.

Optimal layer selection for latent data augmentation

While data augmentation (DA) is generally applied to input data, several studies have reported that applying DA to hidden layers in neural networks, i.e., feature augmentation, can improve performance. However, in previous studies, the layers to which DA is applied have not been carefully considered, often being applied randomly and uniformly or only to a specific layer, leaving room for arbitrariness. Thus, in this study, we investigated the trends of suitable layers for applying DA in various experimental configurations, e.g., training from scratch, transfer learning, various dataset settings, and different models. In addition, to adjust the suitable layers for DA automatically, we propose the adaptive layer selection (AdaLASE) method, which updates the ratio to perform DA for each layer based on the gradient descent method during training. The experimental results obtained on several image classification datasets indicate that the proposed AdaLASE method altered the ratio as expected and achieved high overall test accuracy.

Error Analysis and Optimization of Structural Parameters of Spatial Coordinate Testing System Based on Position-Sensitive Detector.

For the research on real-time accurate testing technology for the explosion point spatial coordinate of munitions, its currently commonly used methods such as acoustic-electric detection or high-speed imaging are limited by the field conditions, response rate, cost, and other factors. In this paper, a method of spatial coordinate testing for the explosion point based on a 2D PSD (position-sensitive detector) intersection is proposed, which has the advantages of a faster response, better real-time performance, and a lower cost. Firstly, a mathematical model of the spatial coordinate testing system was constructed, and an error propagation model for structural parameters was developed. The influence of the position of the optical axes' intersection as well as the azimuth angle and pitch angle on the test accuracy of the system was simulated and analyzed, thus obtaining the distribution and variation trend of the overall error propagation coefficient of the system. Finally, experiments were designed to obtain the test error of the system for validation. The results show that the system test accuracy is high when the azimuth angle is 20°-50°, the overall error propagation coefficient does not exceed 48.80, and the average test error is 56.17 mm. When the pitch angle is -2.5°-2.5°, the system has a higher test accuracy, with the overall error propagation coefficient not exceeding 44.82, and the average test error is 41.87 mm. The test accuracy of the system is higher when the position of the optical axes' intersection is chosen to make sure that explosion points fall in the region of the negative half-axis of the Zw-axis of the world coordinate system, with an overall error propagation coefficient of less than 44.78 and an average test error of 73.38 mm. It is shown that a reasonable selection of system structure parameters can significantly improve the system test accuracy and optimize the system deployment mode under the long-distance field conditions so as to improve the deployment efficiency.

Addressing Imbalanced EEG Data for Improved Microsleep Detection: An ADASYN, FFT and LDA-Based Approach

Microsleep, brief lapses in consciousness lasting less than 15 seconds, are often accompanied by feelings of fatigue and are detectable through a deceleration in electroencephalogram (EEG) signal frequencies. Accurate identification of microsleep is critical for assessing driver alertness and preventing accidents. This paper introduces a novel approach to detecting driver microsleep by leveraging EEG signals and advanced machine learning techniques. The methodology begins with preprocessing raw EEG data to improve quality and balance, utilizing the ADASYN algorithm to address dataset imbalances. After preprocessing, features are extracted using Fast Fourier Transform (FFT), which provides a comprehensive frequency domain analysis of the EEG signals. For classification, Linear Discriminant Analysis (LDA) is employed to effectively distinguish between microsleep events and normal wakefulness based on the extracted features. The proposed framework was rigorously validated using a well-established publicly available EEG dataset, which included recordings from 76 healthy individuals. The validation results revealed a high testing accuracy of 92.71% in detecting microsleep episodes, demonstrating the effectiveness of the proposed approach. These results underscore the potential of combining EEG signal analysis with machine learning models for practical applications in monitoring driver alertness. The framework could significantly enhance driver safety by providing an effective tool for detecting microsleep and thereby reducing the risk of accidents caused by drowsy driving. This research highlights the promising application of advanced signal processing and machine learning techniques in the field of driver alertness monitoring.

A Multi-Dimensional Reverse Auction Mechanism for Volatile Federated Learning in the Mobile Edge Computing Systems

Federated learning (FL) can break the problem of data silos and allow multiple data owners to collaboratively train shared machine learning models without disclosing local data in mobile edge computing. However, how to incentivize these clients to actively participate in training and ensure efficient convergence and high test accuracy of the model has become an important issue. Traditional methods often use a reverse auction framework but ignore the consideration of client volatility. This paper proposes a multi-dimensional reverse auction mechanism (MRATR) that considers the uncertainty of client training time and reputation. First, we introduce reputation to objectively reflect the data quality and training stability of the client. Then, we transform the goal of maximizing social welfare into an optimization problem, which is proven to be NP-hard. Then, we propose a multi-dimensional auction mechanism MRATR that can find the optimal client selection and task allocation strategy considering clients’ volatility and data quality differences. The computational complexity of this mechanism is polynomial, which can promote the rapid convergence of FL task models while ensuring near-optimal social welfare maximization and achieving high test accuracy. Finally, the effectiveness of this mechanism is verified through simulation experiments. Compared with a series of other mechanisms, the MRATR mechanism has faster convergence speed and higher testing accuracy on both the CIFAR-10 and IMAGE-100 datasets.

Fault diagnosis method for distribution network based on improved deep learning

Abstract Due to the complexity of the grid structure, limited fault samples, and difficulty in obtaining them, there is a need to improve the accuracy of fault diagnosis in distribution networks. To address the issue of insufficient training samples in the target domain leading to poor training results, a fault diagnosis method for distribution networks based on improved deep learning is proposed. Firstly, finite element simulation is employed to generate fault data under different operating conditions, obtaining a sufficient number of samples for feature extraction and enhancement. Secondly, a deep convolutional neural network (DCNN) model is constructed, and the model is trained using the sample data. Finally, the model is used for fault diagnosis in distribution networks under different operating conditions. Experimental results demonstrate that the fault diagnosis model achieves high testing accuracy and is capable of detecting fault locations and types with high precision.

An attention-based deep learning for acute lymphoblastic leukemia classification

The bone marrow overproduces immature cells in the malignancy known as Acute Lymphoblastic Leukemia (ALL). In the United States, about 6500 occurrences of ALL are diagnosed each year in both children and adults, comprising nearly 25% of pediatric cancer cases. Recently, many computer-assisted diagnosis (CAD) systems have been proposed to aid hematologists in reducing workload, providing correct results, and managing enormous volumes of data. Traditional CAD systems rely on hematologists’ expertise, specialized features, and subject knowledge. Utilizing early detection of ALL can aid radiologists and doctors in making medical decisions. In this study, Deep Dilated Residual Convolutional Neural Network (DDRNet) is presented for the classification of blood cell images, focusing on eosinophils, lymphocytes, monocytes, and neutrophils. To tackle challenges like vanishing gradients and enhance feature extraction, the model incorporates Deep Residual Dilated Blocks (DRDB) for faster convergence. Conventional residual blocks are strategically placed between layers to preserve original information and extract general feature maps. Global and Local Feature Enhancement Blocks (GLFEB) balance weak contributions from shallow layers for improved feature normalization. The global feature from the initial convolution layer, when combined with GLFEB-processed features, reinforces classification representations. The Tanh function introduces non-linearity. A Channel and Spatial Attention Block (CSAB) is integrated into the neural network to emphasize or minimize specific feature channels, while fully connected layers transform the data. The use of a sigmoid activation function concentrates on relevant features for multiclass lymphoblastic leukemia classification The model was analyzed with Kaggle dataset (16,249 images) categorized into four classes, with a training and testing ratio of 80:20. Experimental results showed that DRDB, GLFEB and CSAB blocks’ feature discrimination ability boosted the DDRNet model F1 score to 0.96 with minimal computational complexity and optimum classification accuracy of 99.86% and 91.98% for training and testing data. The DDRNet model stands out from existing methods due to its high testing accuracy of 91.98%, F1 score of 0.96, minimal computational complexity, and enhanced feature discrimination ability. The strategic combination of these blocks (DRDB, GLFEB, and CSAB) are designed to address specific challenges in the classification process, leading to improved discrimination of features crucial for accurate multi-class blood cell image identification. Their effective integration within the model contributes to the superior performance of DDRNet.

Deep Learning for Early Detection of Diabetic Retinopathy: A Colour Fundus Image-based Approach with VGG16 and VGG19

Diabetic retinopathy (DR) is a prevalent complication of diabetes, posing a significant threat to vision. Early detection and intervention are critical to prevent vision loss. However, manual screening of retinal images for DR is time-consuming and susceptible to human error. This study investigates the application of deep learning models, specifically VGG16 and VGG19, for automated DR diagnosis in Colour Fundus images. We evaluate their performance in classifying images as either Diabetic Retinopathy or No Diabetic Retinopathy. Our results demonstrate high testing accuracy with VGG16 achieving 93.96%, suggesting a promising approach for automated DR screening and aiding healthcare professionals in early and accurate diagnosis. Keywords : Convolutional Neural Network, Diabetic Retinopathy, Visual Geometry Group 16 (VGG16), Visual Geometry Group 19 (VGG19), Transfer Learning, Deep Learning.

Research on Water Finding Methods for Horizontal Wells

The outflow of water from horizontal wells will lead to a decrease in oil production and a decrease in production efficiency. The commonly used water exploration methods have problems such as harmful oil layers, long auxiliary logging time, low efficiency, high cost, and discontinuous testing data. The technology of using crawlers to lower and search for water does not require the cooperation of the operation team. Logging work can be carried out during mining, and real-time monitoring of the production well can be carried out. The instrument is kept in the center of the casing, with high testing accuracy and short logging time, which improves production efficiency. In response to the problem of difficulty in locating the water outlet position using the crawling water finding technology, this article establishes a method for determining the water outlet position of a horizontal well based on the actual working conditions of a domestic oil field using the crawling water finding technology.

Optic Disc Segmentation in Human Retina Images Using a Meta Heuristic Optimization Method and Disease Diagnosis with Deep Learning

Glaucoma is a common eye disease that damages the optic nerve and leads to loss of vision. The disease shows few symptoms in the early stages, making its identification a complex task. To overcome the challenges associated with this task, this study aimed to tackle the localization and segmentation of the optic disc, as well as the classification of glaucoma. For the optic disc segmentation, we propose a novel metaheuristic approach called Grey Wolf Optimization (GWO). Two different approaches are used for glaucoma classification: a one-stage approach, in which the whole image without cropping is used for classification, and a two-stage approach. In the two-stage approach, the optic disc region is detected using the You Only Look Once (YOLO) detection algorithm. Once the optic disc region of interest (ROI) is identified, glaucoma classification is performed using pre-trained convolutional neural networks (CNNs) and vision transformation techniques. In addition, both the one-stage and the two-stage approaches are applied in combination with the pre-trained CNN using the Random Forest algorithm. In segmentation, GWO achieved an average sensitivity of 96.04%, a specificity of 99.58%, an accuracy of 99.39%, a DICE coefficient of 94.15%, and a Jaccard index of 90.4% on the Drishti-GS dataset. For classification, the proposed method achieved remarkable results with a high-test accuracy of 100% and 88.18% for hold-out validation and three-fold cross-validation for the Drishti-GS dataset, and 96.15% and 93.84% for ORIGA with hold-out and five-fold cross-validation, respectively. Comparing the results with previous studies, the proposed CNN model outperforms them. In addition, the use of the Swin transformer shows its effectiveness in classifying glaucoma in different subsets of the data.

LabVIEW-based rotary balance data synchronization acquisition system design

Abstract To solve the problem that many high-precision rotating balance bridge output signals and angular phase signals need to be measured synchronously in wind tunnel force measurement experiments, a rotating balance data synchronization acquisition system is designed based on LabVIEW and NI PXIe acquisition board. LabVIEW graphical programming is applied, combined with STM32F407VET6 processor embedded development of the timing divider, using the equal interval method to obtain the required synchronization trigger signals and use them as the digital trigger signals of the NI PXIe acquisition card. In addition, multi-threaded programming technology is used to develop the upper computer software to realize the storage, display, and analysis of the collected data. The test proved that the system has high testing accuracy, a short development cycle, embedded visualization development is easy to debug, stable operation, and real-time and reliable data synchronization acquisition.

Enhancing EfficientNetv2 with global and efficient channel attention mechanisms for accurate MRI-Based brain tumor classification

The early and accurate diagnosis of brain tumors is critical for effective treatment planning, with Magnetic Resonance Imaging (MRI) serving as a key tool in the non-invasive examination of such conditions. Despite the advancements in Computer-Aided Diagnosis (CADx) systems powered by deep learning, the challenge of accurately classifying brain tumors from MRI scans persists due to the high variability of tumor appearances and the subtlety of early-stage manifestations. This work introduces a novel adaptation of the EfficientNetv2 architecture, enhanced with Global Attention Mechanism (GAM) and Efficient Channel Attention (ECA), aimed at overcoming these hurdles. This enhancement not only amplifies the model’s ability to focus on salient features within complex MRI images but also significantly improves the classification accuracy of brain tumors. Our approach distinguishes itself by meticulously integrating attention mechanisms that systematically enhance feature extraction, thereby achieving superior performance in detecting a broad spectrum of brain tumors. Demonstrated through extensive experiments on a large public dataset, our model achieves an exceptional high-test accuracy of 99.76%, setting a new benchmark in MRI-based brain tumor classification. Moreover, the incorporation of Grad-CAM visualization techniques sheds light on the model’s decision-making process, offering transparent and interpretable insights that are invaluable for clinical assessment. By addressing the limitations inherent in previous models, this study not only advances the field of medical imaging analysis but also highlights the pivotal role of attention mechanisms in enhancing the interpretability and accuracy of deep learning models for brain tumor diagnosis. This research sets the stage for advanced CADx systems, enhancing patient care and treatment outcomes.

SFL-TUM: Energy efficient SFRL method for large scale AI model's task offloading in UAV-assisted MEC networks

The convergence of mobile edge computing (MEC) network with unmanned aerial vehicles (UAVs) presents an auspicious opportunity to revolutionize wireless communication and facilitate high-speed internet access in remote regions for mobile devices (MDs) as well as large scale artificial intelligence (AI) models. However, the substantial amount of data produced by the UAVs-assisted MEC network necessitates the integration of efficient distributed learning techniques in AI models. In recent times, distributed learning algorithms, including federated reinforcement learning (FRL) and split learning (SL), have been explored for the purpose of learning machine learning (ML) models that are distributed by sharing model parameters, as opposed to large raw data-sets as seen in traditional centralized learning algorithms. To implement the hybrid method, the model is first trained locally on each UAV-assisted MEC network using SL. Subsequently, the model parameters that have been encrypted are sent to a central server for federated averaging. Finally, after the model has been updated, it is distributed to each UAV-assisted MEC network for local fine-tuning. Our simulations indicate that the proposed split and federated reinforcement learning (SFRL) framework yields comparable high-test accuracy performance while consuming less energy compared to extant distributed learning algorithms. Furthermore, the SFRL algorithm efficiently realizes energy-efficient selection between the SL and FRL methods under different distributions. Numerical results shows that the proposed scheme improves the accuracy by 29.31% and reduced the energy consumption by around 67.34% and time delay by about 7.37%. as compared to the existing baseline schemes.

Transformer based ensemble deep learning approach for remote sensing natural scene classification

ABSTRACT Very high resolution (VHR) remote sensing (RS) image classification is paramount for detailed Earth’s surface analysis. Feature extraction from VHR natural scenes is crucial, but it becomes a challenging task because of the overlapping edges present in images. Multiple open-source datasets exist in the literature to train robust models, and they have been benchmarked using deep learning models. However, different datasets contain different numbers of classes; a few of them could be absent in other datasets because they are independent of the other classes, and the classes are sometimes not mutually exclusive amongst the same dataset. Thus, it is very challenging to generalize a model trained on a single dataset to perform scene classification on unknown classes of multiple benchmark datasets and real-time images. Thus, this work introduces the Remote Sensing Natural Scenes 92 (RS_NS92) dataset, consisting of 36,785 images belonging to 92 classes, curated by selectively taking the union of all subclasses from five benchmark datasets. This class count is significantly higher than publicly available datasets and maintains a low-class imbalance and a comprehensive data distribution for robust model training. It also provides the remote sensing community with an extra platform to validate the performance on multiple benchmarks. Inspired by federated learning, an ensemble approach consisting of three feature extraction backbones: Vision Transformers, Swin Transformers, and ConvNeXt (termed the VSC_Ensemble model) is also introduced. This model can make extraordinary predictions across multiple datasets by finetuning weights using transfer learning. Experimental analysis with the proposed approach not only obtains a high test accuracy of 97.24% and an F1-score of 0.9587 for the 92 classes on a 90:10 split of the proposed benchmark dataset but also gets excellent results on unseen test images of other datasets, which are comparable to the state-of-the-art results.

Addressing equifinality in agent-based modeling: a sequential parameter space search method based on sensitivity analysis

This study addresses the challenge of equifinality in agent-based modeling (ABM) by introducing a novel sequential calibration approach. Equifinality arises when multiple models equally fit observed data, risking the selection of an inaccurate model. In the context of ABM, such a situation might arise due to limitations in data, such as aggregating observations into coarse spatial units. It can lead to situations where successfully calibrated model parameters may still result in reliability issues due to uncertainties in accurately calibrating the inner mechanisms. To tackle this, we propose a method that sequentially calibrates model parameters using diverse outcomes from multiple datasets. The method aims to identify optimal parameter combinations while mitigating computational intensity. We validate our approach through indoor pedestrian movement simulation, utilizing three distinct outcomes: (1) the count of grid cells crossed by individuals, (2) the number of people in each grid cell over time (fine grid) and (3) the number of people in each grid cell over time (coarse grid). As a result, the optimal calibrated parameter combinations were selected based on high test accuracy to avoid overfitting. This method addresses equifinality while reducing computational intensity of parameter calibration for spatially explicit models, as well as ABM in general.

A Novel Deep Convolutional Neural Network Approach using Jacobi Polynomial and Laplacian Function (JPLF) in Recognition of Plant Leaf Disease

Background/Objectives: Enhancing agricultural productivity is crucial for fostering economic growth. Plant diseases significantly threaten crops, necessitating timely detection to mitigate adverse impacts on quality, quantity, and overall productivity. This research addresses the importance of early disease detection in agriculture and proposes an innovative method utilizing Jacobian Polynomial and Laplacian Function for precise identification. Methods: Efficient monitoring of large-scale crop farms with minimal workforce is essential. To achieve this, an automatic method for plant disease detection is proposed. The method leverages Jacobian polynomials to expand input features, mitigating correlation issues among input vectors. The expanded Jacobi polynomial is the input vector for a backpropagation algorithm with a novel activation function based on the Laplacian function. Findings: The efficacy of the proposed JPLF model is demonstrated through the accurate identification of leaf diseases, achieving a high testing accuracy of 92.07%. Comparative analysis with existing models, such as CNN with MobileNet V2 (85.38%) and the IoU model (83.75%), highlights the superiority of the JPLF model in plant disease detection. Novelty: To overcome the limitations of existing approaches, the incorporation of Jacobian polynomials plays a pivotal role in expanding input features. This expansion aids in eliminating correlations among input vectors, enhancing the efficacy of disease detection. The proposed model, Jacobi Polynomial and Laplacian Function (JPLF) introduces a unique activation function based on the Laplacian function, improving accuracy. Keywords: Plant Disease Detection, Jacobi Polynomial, Laplacian Transform, Deep Learning Model, Feature Expansion

Enhancing Federated Learning With Server-Side Unlabeled Data by Adaptive Client and Data Selection

Federated learning (FL) has been widely applied to collaboratively train deep learning (DL) models on massive end devices (i.e., clients). Due to the limited storage capacity and high labeling cost, the data on each client may be insufficient for model training. Conversely, in cloud datacenters, there exist large-scale unlabeled data, which are easy to collect from public access (e.g., social media). Herein, we propose the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ada-FedSemi</i> system, which leverages both on-device labeled data and in-cloud unlabeled data to boost the performance of DL models. In each round, local models are aggregated to produce pseudo-labels for the unlabeled data, which are utilized to enhance the global model. Considering that the number of participating clients and the quality of pseudo-labels will have a significant impact on the training performance, we introduce a multi-armed bandit (MAB) based online algorithm to adaptively determine the participating fraction and confidence threshold. Besides, to alleviate the impact of stragglers, we assign local models of different depths for heterogeneous clients. Extensive experiments on benchmark models and datasets show that given the same resource budget, the model trained by Ada-FedSemi achieves 3% <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\sim$</tex-math></inline-formula> 14.8% higher test accuracy than that of the baseline methods. When achieving the same test accuracy, Ada-FedSemi saves up to 48% training cost, compared with the baselines. Under the scenario with heterogeneous clients, the proposed HeteroAda-FedSemi can further speed up the training process by 1.3× <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\sim 1.5\times$</tex-math></inline-formula> .

Classification of Childhood Obstructive Sleep Apnea based on X-ray images analysis by Quasi-conformal Geometry

Craniofacial profile is one of the anatomical causes of obstructive sleep apnea (OSA). By medical research, cephalometry provides information on patients’ skeletal structures and soft tissues. In this work, a novel approach to cephalometric analysis using quasi-conformal geometry based local deformation information was proposed for OSA classification. Our study was a retrospective analysis based on 60 case-control pairs with accessible lateral cephalometry and polysomnography (PSG) data (mean age: 8.9 ± 2.3 years). By using the quasi-conformal geometry to study the local deformation around 15 landmark points, and combining the results with three linear distances between landmark points, a total of 1218 information features were obtained per subject. A L2 norm based classification model was built. Under experiments, our proposed model achieves high testing accuracy. The accurate classification tool provides us with a reliable and accurate screening tool for childhood OSA.