Use Of Deep Learning Models Research Articles

UAV4TREE: DEEP LEARNING-BASED SYSTEM FOR AUTOMATIC CLASSIFICATION OF TREE SPECIES USING RGB OPTICAL IMAGES OBTAINED BY AN UNMANNED AERIAL VEHICLE

Abstract. Automated tree classification from unmanned aerial vehicle (UAV) images is a challenging task with several applications in forest management and conservation. In this study, we propose UAV4Tree a Deep Learning based system that automatically classifies RGB optical images obtained by the UAV. In particular, we explore the use of augmented datasets and various deep learning models, including ResNet, DenseNet, InceptionV3, and Vision Transformer, for the classification of tree images obtained from UAVs. Our experiments show that the use of an augmented dataset can significantly improve the accuracy of the classification by approximately 10 points compared to the use of a non-augmented dataset. We also found that fine-tuning and the introduction of dropout were essential for improving the generalization ability of the models on the augmented dataset. Furthermore, the use of Super Resolution Generative Adversarial Network (SR-GAN) in the original dataset allowed us to increase the performance of some models. Our findings provide valuable insights into the use of deep learning models for automated tree classification from UAV imagery, which has significant implications for sustainable forest management and conservation.

GA-based weighted ensemble learning for multi-label aerial image classification using convolutional neural networks and vision transformers

Multi-label classification (MLC) of aerial images is a crucial task in remote sensing image analysis. Traditional image classification methods have limitations in image feature extraction, leading to an increasing use of deep learning models, such as convolutional neural networks (CNN) and vision transformers (ViT). However, the standalone use of these models may have limitations when dealing with MLC. To enhance the generalization performance of MLC of aerial images, this paper combines two CNN and two ViT models, comparing four single deep learning models, a manually weighted ensemble learning method, and a GA-based weighted ensemble method. The experimental results using two public multi-label aerial image datasets show that the classification performance of ViT models is better than CNN models, the traditional weighted ensemble learning model performs better than a single deep learning model, and the GA-based weighted ensemble method performs better than the manually weighted ensemble learning method. The GA-based weighted ensemble method proposed in this study can achieve better MLC performance of aerial images than previous results.

Use of deep learning model for paediatric elbow radiograph binomial classification: initial experience, performance and lessons learnt

Abstract Introduction: In this study, we aimed to compare the performance of a convolutional neural network (CNN)-based deep learning model that was trained on a dataset of normal and abnormal paediatric elbow radiographs with that of paediatric emergency department (ED) physicians on a binomial classification task. Methods: A total of 1,314 paediatric elbow lateral radiographs (patient mean age 8.2 years) were retrospectively retrieved and classified based on annotation as normal or abnormal (with pathology). They were then randomly partitioned to a development set (993 images); first and second tuning (validation) sets (109 and 100 images, respectively); and a test set (112 images). An artificial intelligence (AI) model was trained on the development set using the EfficientNet B1 network architecture. Its performance on the test set was compared to that of five physicians (inter-rater agreement: fair). Performance of the AI model and the physician group was tested using McNemar test. Results: The accuracy of the AI model on the test set was 80.4% (95% confidence interval [CI] 71.8%–87.3%), and the area under the receiver operating characteristic curve (AUROC) was 0.872 (95% CI 0.831–0.947). The performance of the AI model vs. the physician group on the test set was: sensitivity 79.0% (95% CI: 68.4%–89.5%) vs. 64.9% (95% CI: 52.5%–77.3%; P = 0.088); and specificity 81.8% (95% CI: 71.6%–92.0%) vs. 87.3% (95% CI: 78.5%–96.1%; P = 0.439). Conclusion: The AI model showed good AUROC values and higher sensitivity, with the P-value at nominal significance when compared to the clinician group.

Deep Random Forest and AraBert for Hate Speech Detection from Arabic Tweets

Nowadays, hate speech detection from Arabic tweets attracts the attention of many researchers. Numerous systems and techniques have been proposed to address this classification challenge. Nonetheless, three major limits persist: the use of deep learning models with an excess of hyperparameters, the reliance on hand-crafted features, and the requirement for a huge amount of training data to achieve satisfactory performance. In this study, we propose Contextual Deep Random Forest (CDRF), a hate speech detection approach that combines contextual embedding and Deep Random Forest. From the experimental findings, the Arabic contextual embedding model proves to be highly effective in hate speech detection, outperforming the static embedding models. Additionally, we prove that the proposed CDRF significantly enhances the performance of Arabic hate speech classification.

Deep Learning and Statistical Models for Forecasting Transportation Demand: A Case Study of Multiple Distribution Centers

Background: Transportation demand forecasting is an essential activity for logistics operators and carriers. It leverages business operation decisions, infrastructure, management, and resource planning activities. Since 2015, there has been an increase in the use of deep learning models in this domain. However, there is a gap in works comparing traditional statistics and deep learning models for transportation demand forecasts. This work aimed to perform a case study of aggregated transportation demand forecasts in 54 distribution centers of a Brazilian carrier. Methods: A computational simulation and case study methods were applied, exploring the characteristics of the datasets through autoregressive integrated moving average (ARIMA) and its variations, in addition to a deep neural network, long short-term memory, known as LSTM. Eight scenarios were explored while considering different data preprocessing methods and evaluating how outliers, training and testing dataset splits during cross-validation, and the relevant hyperparameters of each model can affect the demand forecast. Results: The long short-term memory networks were observed to outperform the statistical methods in ninety-four percent of the dispatching units over the evaluated scenarios, while the autoregressive integrated moving average modeled the remaining five percent. Conclusions: This work found that forecasting transportation demands can address practical issues in supply chains, specially resource planning management.

Deep learning framework for epidemiological forecasting: A study on COVID-19 cases and deaths in the Amazon state of Pará, Brazil.

Modeling time series has been a particularly challenging aspect due to the need for constant adjustments in a rapidly changing environment, data uncertainty, dependencies between variables, volatile fluctuations, and the need to identify ideal hyperparameters. The present study presents a Framework capable of making projections from time series related to cases and deaths by COVID-19 in the Amazonian state of Pará, in Brazil. For the first time, deep learning models such as TCN, TRANSFORMER, TFT, N-BEATS, and N-HiTS were assessed for this purpose. The ARIMA statistical model was also used in post-processing for residual adjustment and short-term smoothing of the generated forecasts. The Framework generates probabilistic forecasts, with multivariate support, considering the following variables: daily cases per day of the first symptom, cases published daily, the occurrence of deaths, deaths published daily, and percentage of daily vaccination. The generated predictions are statistically evaluated by determining the best model for 7-day moving average projections using evaluating metrics such as MSE, RMSE, MAPE, sMAPE, r2, Coefficient of Variation, and residual analysis. As a result, the generated projections showed an average error of 5.4% for Cases Publication, 8.0% for Cases Symptoms, 11.12% for Deaths Publication, and 4.6% for Deaths Occurrence, with the N-HiTS and N-BEATS models obtaining better results. In general terms, the use of deep learning models to predict cases and deaths from COVID-19 has proven to be a valuable practice for analyzing the spread of the virus, which allows health managers to better understand and respond to this kind of pandemic outbreak.

Identification of runner fatigue stages based on inertial sensors and deep learning.

Introduction: Running is one of the most popular sports in the world, but it also increases the risk of injury. The purpose of this study was to establish a modeling approach for IMU-based subdivided action pattern evaluation and to investigate the classification performance of different deep models for predicting running fatigue. Methods: Nineteen healthy male runners were recruited for this study, and the raw time series data were recorded during the pre-fatigue, mid-fatigue, and post-fatigue states during running to construct a running fatigue dataset based on multiple IMUs. In addition to the IMU time series data, each participant's training level was monitored as an indicator of their level of physical fatigue. Results: The dataset was examined using single-layer LSTM (S_LSTM), CNN, dual-layer LSTM (D_LSTM), single-layer LSTM plus attention model (LSTM + Attention), CNN, and LSTM hybrid model (LSTM + CNN) to classify running fatigue and fatigue levels. Discussion: Based on this dataset, this study proposes a deep learning model with constant length interception of the raw IMU data as input. The use of deep learning models can achieve good classification results for runner fatigue recognition. Both CNN and LSTM can effectively complete the classification of fatigue IMU data, the attention mechanism can effectively improve the processing efficiency of LSTM on the raw IMU data, and the hybrid model of CNN and LSTM is superior to the independent model, which can better extract the features of raw IMU data for fatigue classification. This study will provide some reference for many future action pattern studies based on deep learning.

Deep Learning Model Transfer in Forest Mapping Using Multi-Source Satellite SAR and Optical Images

Deep learning (DL) models are gaining popularity in forest variable prediction using Earth observation (EO) images. However, in practical forest inventories, reference datasets are often represented by plot- or stand-level measurements, while high-quality representative wall-to-wall reference data for end-to-end training of DL models are rarely available. Transfer learning facilitates expansion of the use of deep learning models into areas with sub-optimal training data by allowing pretraining of the model in areas where high-quality teaching data are available. In this study, we perform a “model transfer” (or domain adaptation) of a pretrained DL model into a target area using plot-level measurements and compare performance versus other machine learning models. We use an earlier developed UNet based model (SeUNet) to demonstrate the approach on two distinct taiga sites with varying forest structure and composition. The examined SeUNet model uses multi-source EO data to predict forest height. Here, EO data are represented by a combination of Copernicus Sentinel-1 C-band SAR and Sentinel-2 multispectral images, ALOS-2 PALSAR-2 SAR mosaics and TanDEM-X bistatic interferometric radar data. The training study site is located in Finnish Lapland, while the target site is located in Southern Finland. By leveraging transfer learning, the SeUNet prediction achieved root mean squared error (RMSE) of 2.70 m and R2 of 0.882, considerably more accurate than traditional benchmark methods. We expect such forest-specific DL model transfer can be suitable also for other forest variables and other EO data sources that are sensitive to forest structure.

A Review on Speech Recognition for Under-Resourced Languages

Fundamental speech recognition technologies for high-resourced languages are currently successful to build high-quality applications with the use of deep learning models. However, the problem of “borrowing” these speech recognition technologies for under-resourced languages like Vietnamese still has challenges. This study reviews fundamental studies on speech recognition in general as well as speech recognition in Vietnamese, an under-resourced language in particular. Then, it specifies the urgent issues that need current research attention to build Vietnamese speech recognition applications in practice, especially the need to build an open large sentence-labeled speech corpus and open platform for related research, which mostly benefits small individuals/organizations who do not have enough resources.

BERT-Based Approaches to Identifying Malicious URLs.

Malicious uniform resource locators (URLs) are prevalent in cyberattacks, particularly in phishing attempts aimed at stealing sensitive information or distributing malware. Therefore, it is of paramount importance to accurately detect malicious URLs. Prior research has explored the use of deep-learning models to identify malicious URLs, using the segmentation of URL strings into character-level or word-level tokens, and embedding and employing trained models to differentiate between URLs. In this study, a bidirectional encoder representation from a transformers-based (BERT) model was devised to tokenize URL strings, employing its self-attention mechanism to enhance the understanding of correlations among tokens. Subsequently, a classifier was employed to determine whether a given URL was malicious. In evaluating the proposed methods, three different types of public datasets were utilized: a dataset consisting solely of URL strings from Kaggle, a dataset containing only URL features from GitHub, and a dataset including both types of data from the University of New Brunswick, namely, ISCX 2016. The proposed system achieved accuracy rates of 98.78%, 96.71%, and 99.98% on the three datasets, respectively. Additionally, experiments were conducted on two datasets from different domains-the Internet of Things (IoT) and Domain Name System over HTTPS (DoH)-to demonstrate the versatility of the proposed model.

An Intelligent Prairie Dog Optimization (IPDO) and Deep Auto-Neural Network (DANN) based IDS for WSN Security

Wireless sensor networks (WSNs) are targets of intrusion, which seeks to make these networks less capable of performing their duties or even completely eradicate them. The Intrusion Detection System (IDS) is highly important for WSN, since it aids in the identification and detection of harmful attacks that impair the network's regular functionality. In order to strengthen the security of WSN, several machine learning and deep learning approaches are employed in the traditional works. However, its main drawbacks are computational burden, system complexity, poor network performance outcomes, and high false alarms. Therefore, the goal of this study is to develop an intelligent IDS framework for significantly enhancing WSN security through the use of deep learning model. Here, the min-max normalization and data discretization operations are carried out to produce the preprocessed dataset. Then, an Intelligent Prairie Dog Optimization (IPDO) algorithm is used to reduce the dimensionality of features by identifying the best optimal solution with a higher convergence rate. Moreover, a Deep Auto-Neural Network (DANN) based classification method is used to properly forecast the class of data with less false alarms and higher detection rate. For evaluation, a thorough analysis is conducted to evaluate the performance and detection results of the proposed IPDO-DANN model.

Interpretable deep learning for hourly solar radiation prediction: A real measured data case study in Tokyo

With their superior performance in capturing complex relationships between input and output variables, deep learning models have gained popularity in the field of solar radiation prediction. However, the use of deep learning models as a predictive tool is often constrained by their inherent complexity, which can limit the interpretability and usability of the model in practical scenarios, such as model predictive control (MPC) in energy systems. In this study, time series are transformed into an image, and the hourly solar radiation prediction is realized through a convolutional neural network (CNN). The experiments demonstrate that the proposed CNN model enhances interpretability at the cost of a minor 6 % reduction in prediction accuracy, compared to the baseline model. Analogous to the deconvolution operation for image classification, feature maps of CNN networks can thus be used for model interpretability, whereby the closer the feature maps are to the input (In the context of forward propagation, this means that the layers of a CNN closer to the input data), the higher is the interpretability. In extracting the feature map using CNN, solar radiation and radiation time played the most important roles in the entire prediction process.

Construction of Asbestos Slate Deep-Learning Training-Data Model Based on Drone Images.

The detection of asbestos roof slate by drone is necessary to avoid the safety risks and costs associated with visual inspection. Moreover, the use of deep-learning models increases the speed as well as reduces the cost of analyzing the images provided by the drone. In this study, we developed a comprehensive learning model using supervised and unsupervised classification techniques for the accurate classification of roof slate. We ensured the accuracy of our model using a low altitude of 100 m, which led to a ground sampling distance of 3 cm/pixel. Furthermore, we ensured that the model was comprehensive by including images captured under a variety of light and meteorological conditions and from a variety of angles. After applying the two classification methods to develop the learning dataset and employing the as-developed model for classification, 12 images were misclassified out of 475. Visual inspection and an adjustment of the classification system were performed, and the model was updated to precisely classify all 475 images. These results show that supervised and unsupervised classification can be used together to improve the accuracy of a deep-learning model for the detection of asbestos roof slate.

A novel deep learning-based approach for prediction of neonatal respiratory disorders from chest X-ray images

In recent years, many diseases can be diagnosed in a short time with the use of deep learning models in the field of medicine. Most of the studies in this area focus on adult or pediatric patients. However, deep learning studies for the diagnosis of diseases in neonatal are not sufficient. Also, since it is known that respiratory disorders such as pneumonia have a large place among the causes of neonatal death, early and accurate diagnosis of respiratory diseases in neonates is crucial. For this reason, our study aims to detect the presence of respiratory disorders through the developed deep-learning approach using chest X-ray images of patients hospitalized in the Neonatal Intensive Care Unit. Accordingly, the enhanced version of C+EffxNet, the new hybrid deep learning model, is designed to predict respiratory disorders in neonates. In this version, the features selected by PCA are combined as 100, 200, and 300, then the binary classification process was carried out. In the study, the accuracy and kappa value were obtained as 0.965, and 0.904, respectively before feature merging, while these values were obtained as 0.977, and 0.935 after feature merging. This method, which was developed for the diagnosis of respiratory disorders in neonates, was also subsequently applied to a chest X-ray dataset that is frequently used in the literature for the diagnosis of pediatric pneumonia. For this data set, while the accuracy was 0.992, the kappa value was 0.982. The results obtained confirm the success of the proposed method for both datasets.

Deep and Transfer Learning Approaches for Automated Early Detection of Monkeypox (Mpox) Alongside Other Similar Skin Lesions and Their Classification.

The world faces multiple public health emergencies simultaneously, such as COVID-19 and Monkeypox (mpox). mpox, from being a neglected disease, has emerged as a global threat that has spread to more than 100 nonendemic countries, even as COVID-19 has been spreading for more than 3 years now. The general mpox symptoms are similar to chickenpox and measles, thus leading to a possible misdiagnosis. This study aimed at facilitating a rapid and high-brevity mpox diagnosis. Reportedly, mpox circulates among particular groups, such as sexually promiscuous gay and bisexuals. Hence, selectively vaccinating, isolating, and treating them seems difficult due to the associated social stigma. Deep learning (DL) has great promise in image-based diagnosis and could help in error-free bulk diagnosis. The novelty proposed, the system adopted, and the methods and approaches are discussed in the article. The present work proposes the use of DL models for automated early mpox diagnosis. The performances of the proposed algorithms were evaluated using the data set available in public domain. The data set adopted for the study was meant for both training and testing, the details of which are elaborated. The performances of CNN, VGG19, ResNet 50, Inception v3, and Autoencoder algorithms were compared. It was concluded that CNN, VGG19, and Inception v3 could help in early detection of mpox skin lesions, and Inception v3 returned the best (96.56%) classification accuracy.

The use of deep learning models to predict progression-free survival in patients with neuroendocrine tumors.

Aim: The RAISE project assessed whether deep learning could improve early progression-free survival (PFS) prediction in patients with neuroendocrine tumors. Patients & methods: Deep learning models extracted features from CT scans from patients in CLARINET (NCT00353496) (n=138/204). A Cox model assessed PFS prediction when combining deep learning with the sum of longest diameter ratio (SLDr) and logarithmically transformed CgA concentration (logCgA), versus SLDr and logCgA alone. Results: Deep learning models extracted features other than lesion shape to predict PFS at week 72. No increase in performance was achieved with deep learning versus SLDr and logCgA models alone. Conclusion: Deep learning models extracted relevant features to predict PFS, but did not improve early prediction based on SLDr and logCgA.

Deep Learning for Medical Image Cryptography: A Comprehensive Review

Electronic health records (EHRs) security is a critical challenge in the implementation and administration of Internet of Medical Things (IoMT) systems within the healthcare sector’s heterogeneous environment. As digital transformation continues to advance, ensuring privacy, integrity, and availability of EHRs become increasingly complex. Various imaging modalities, including PET, MRI, ultrasonography, CT, and X-ray imaging, play vital roles in medical diagnosis, allowing healthcare professionals to visualize and assess the internal structures, functions, and abnormalities within the human body. These diagnostic images are typically stored, shared, and processed for various purposes, including segmentation, feature selection, and image denoising. Cryptography techniques offer a promising solution for protecting sensitive medical image data during storage and transmission. Deep learning has the potential to revolutionize cryptography techniques for securing medical images. This paper explores the application of deep learning techniques in medical image cryptography, aiming to enhance the privacy and security of healthcare data. It investigates the use of deep learning models for image encryption, image resolution enhancement, detection and classification, encrypted compression, key generation, and end-to-end encryption. Finally, we provide insights into the current research challenges and promising directions for future research in the field of deep learning applications in medical image cryptography.

A novel data augmentation approach for enhancement of ECG signal classification

The use of deep learning models in the classification of medical diseases has evolved drastically in recent years. One such prominent application was in the classification of ECG data with the goal of achieving high accuracy and computational efficiency. However, several challenges remain to be resolved to achieve this goal, such as finding a generalized model, data imbalance, and preserving patient privacy. This research aims to overcome the first two challenges. The novelty of the proposed work is to introduce unique data augmentation for time series ECG signal data based on simple segmentation and then re-arrangement of them. Simple in nature makes it more convenient to use minimal memory for the processing and allows to obtain well-distinguished synthetic signals. Additionally, a less computationally expensive four layers convolutional neural network (CNN) model has also been proposed. The original ECG signal wrapped in a Numpy array was divided into segments of identical length. During segment re-organization, a few section values are investigated, crucial for joining the segments back into a proper signal structure. The proposed algorithm has been evaluated in two ways, including similarity & features measure and classification accuracy through customized CNN & transfer learning model with five-fold cross-validation. The study results revealed that the proposed augmentation algorithm achieved a validation accuracy of 89.87% with 88.99% recall and 0.291 loss by the four layers CNN model. The presented work can be implemented in two ways, i.e., ECG signals dataset enhancement and better utilization of computational resources through the simple CNN model architecture.

Use of artificial intelligence and deep learning in fetal ultrasound imaging.

Deep learning is considered the leading artificial intelligence tool in image analysis in general. Deep-learning algorithms excel at image recognition, which makes them valuable in medical imaging. Obstetric ultrasound has become the gold standard imaging modality for detection and diagnosis of fetal malformations. However, ultrasound relies heavily on the operator's experience, making it unreliable in inexperienced hands. Several studies have proposed the use of deep-learning models as a tool to support sonographers, in an attempt to overcome these problems inherent to ultrasound. Deep learning has many clinical applications in the field of fetal imaging, including identification of normal and abnormal fetal anatomy and measurement of fetal biometry. In this Review, we provide a comprehensive explanation of the fundamentals of deep learning in fetal imaging, with particular focus on its clinical applicability. © 2022 International Society of Ultrasound in Obstetrics and Gynecology.

Investigation and benchmarking of U-Nets on prostate segmentation tasks.

In healthcare, a growing number of physicians and support staff are striving to facilitate personalized radiotherapy regimens for patients with prostate cancer. This is because individual patient biology is unique, and employing a single approach for all is inefficient. A crucial step for customizing radiotherapy planning and gaining fundamental information about the disease, is the identification and delineation of targeted structures. However, accurate biomedical image segmentation is time-consuming, requires considerable experience and is prone to observer variability. In the past decade, the use of deep learning models has significantly increased in the field of medical image segmentation. At present, a vast number of anatomical structures can be demarcated on a clinician's level with deep learning models. These models would not only unload work, but they can offer unbiased characterization of the disease. The main architectures used in segmentation are the U-Net and its variants, that exhibit outstanding performances. However, reproducing results or directly comparing methods is often limited by closed source of data and the large heterogeneity among medical images. With this in mind, our intention is to provide a reliable source for assessing deep learning models. As an example, we chose the challenging task of delineating the prostate gland in multi-modal images. First, this paper provides a comprehensive review of current state-of-the-art convolutional neural networks for 3D prostate segmentation. Second, utilizing public and in-house CT and MR datasets of varying properties, we created a framework for an objective comparison of automatic prostate segmentation algorithms. The framework was used for rigorous evaluations of the models, highlighting their strengths and weaknesses.