Search

In recent years, the advancement of convolutional neural networks (CNNs) has been driven by the pursuit of higher classification accuracy in image tasks. However, achieving optimal performance often requires extensive manual design, incorporating domain-specific knowledge and problem-understanding. This approach often results in highly complex network architectures, overlooking the potential drawbacks of such complexity. To this end, we propose MOGA-CNN, a Multi-Objective Genetic Algorithm for CNN structure that treats the CNN architecture design as a bi-objective optimization problem. MOGA-CNN aims to simultaneously optimize classification accuracy and minimize computational complexity, as measured by the number of learnable parameters. We employ the NSGA-II algorithm to effectively explore the trade-offs between these two conflicting objectives. The main contribution of this paper is the development of an encoding mechanism that captures the essential hyperparameters that influence CNN architecture, including the fully connected layer. To evaluate the effectiveness of our proposed algorithm, we conducted extensive experiments on four datasets, comparing its performance against other state-of-the-art methods. The results consistently demonstrate that our approach achieves satisfactory results when compared to these approaches. [ABSTRACT FROM AUTHOR]

In modern healthcare, medical image processing plays a vital role in enabling early disease detection, treatment planning, and improved patient care. However, traditional methods face challenges such as handling big data, scalability, and computational intensity. To address these issues, this paper proposes a Convolutional Global Gated Recurrent-based Adaptive Gazelle (CGGR-AG) algorithm for medical image processing applications. The CGGR-AG algorithm detects abnormalities and classifies specific objects within images by leveraging Convolutional Neural Networks (CNNs) for feature extraction and Gated Recurrent Units (GRUs) for capturing sequential patterns. Additionally, the Adaptive Gazelle Optimization algorithm fine-tunes parameters to enhance the effectiveness of the CGGR-AG method. Experimental validation is conducted on Tuberculosis and heart disease datasets, evaluating performance metrics including recall, specificity, accuracy, Area Under the Curve – Receiver Operating Characteristic (AUC-ROC), precision, and F1-score. Comparative analysis with state-of-the-art methods demonstrates the effectiveness of the CGGR-AG method in medical image processing applications. [ABSTRACT FROM AUTHOR]

The study and development of fault detection and diagnosis (FDD) systems are relevant tasks for industrial processes. Another prominent field is applying deep learning (DL) models to solve engineering problems, such as FDD systems' design. Often, the preliminary tests are conducted using simulated datasets to verify the chosen methodology and avoid unnecessarily disturbing the real process. Even if the data used come from a computer simulation, it must remain as realistic as possible. In several studies, researchers have used the Tennessee Eastman Process (TEP) benchmark for addressing the application of DL models to build effective FDD frameworks. However, most of them use preexisting datasets, and this presents some drawbacks that can negatively impact the DL model's training stage. In addition, none of them have evaluated how to adjust the existing FDD model when the process control strategy is changed. This paper presents various topologies of convolutional neural networks (CNNs) to model a FDD system for the TEP benchmark using new datasets. For the first time, we investigate the performance of fully convolutional networks (FCNs) in the TEP study case. Additionally, we apply transfer learning (TL) to surpass the model inadequacy when the data distribution changes due to an alteration in the process' closed-loop system. [ABSTRACT FROM AUTHOR]

Nucleus segmentation is an imperative step in the qualitative study of imaging datasets, considered as an intricate task in histopathology image analysis. Segmenting a nucleus is an important part of diagnosing, staging, and grading cancer, but overlapping regions make it hard to separate and tell apart independent nuclei. Deep Learning is swiftly paving its way in the arena of nucleus segmentation, attracting quite a few researchers with its numerous published research articles indicating its efficacy in the field. This paper presents a systematic survey on nucleus segmentation using deep learning in the last five years (2017–2021), highlighting various segmentation models (U-Net, SCPP-Net, Sharp U-Net, and LiverNet) and exploring their similarities, strengths, datasets utilized, and unfolding research areas. [ABSTRACT FROM AUTHOR]

COVID-19 is an acronym for coronavirus disease 2019. Initially, it was called 2019-nCoV, and later International Committee on Taxonomy of Viruses (ICTV) termed it SARS-CoV-2. On 30th January 2020, the World Health Organization (WHO) declared it a pandemic. With an increasing number of COVID-19 cases, the available medical infrastructure is essential to detect the suspected cases. Medical imaging techniques such as Computed Tomography (CT), chest radiography can play an important role in the early screening and detection of COVID-19 cases. It is important to identify and separate the cases to stop the further spread of the virus. Artificial Intelligence can play an important role in COVID-19 detection and decreases the workload on collapsing medical infrastructure. In this paper, a deep convolutional neural network-based architecture is proposed for the COVID-19 detection using chest radiographs. The dataset used to train and test the model is available on different public repositories. Despite having the high accuracy of the model, the decision on COVID-19 should be made in consultation with the trained medical clinician. [ABSTRACT FROM AUTHOR]

Recognizing a person's physical activity with certainty makes an important aspect of intelligent computing. Modern smart devices are equipped with powerful sensors that are suitable for sensor-based human activity recognition (AR) task. Traditional approaches to human activity recognition has made significant progress but most of those methods rely upon manual feature extraction. The design and selection of relevant features is the most challenging task in sensor-based human AR problem. Using manually extracted features for this task hinders the generalization of performance and these handcrafted features are also incapable of handling similar and complex activities with certainty. In this paper, we propose a deep learning based method for human activity recognition problem. The method uses convolutional neural networks to automatically extract features from raw sensor data and classify six basic human activities. Furthermore, transfer learning is used to reduce the computational cost involved in training the model from scratch for a new user. The model uses the labelled information from supervised learning, to mutually enhance the feature extraction and classification. Experiments carried on benchmark dataset verified the strong advantage of proposed method over the traditional human AR algorithms such as Random Forest (RF) and multiclass Support Vector Machine (SVM). [ABSTRACT FROM AUTHOR]

Currently, the most widespread infectious illness in the world is the coronavirus (COVID-19). The original diagnosis of this illness presents the most obstacle in preventing subsequent infections and their transmission from one person to another. Therefore, it is crucial to employ both a clinical process and an automated diagnostic technology for the quick detection of COVID-19 to stop its spread. Chest X-ray (CXR) images from chest radiography could be used in artificial intelligence (AI) approaches to diagnose COVID-19 with excellent diagnostic accuracy. In this research, a new support vector machine kernel (SVM Kernel) and convolutional neural network (CNN) combination is suggested to classify COVID-19 using X-ray images. The fact that there are relatively few studies in the literature that provide novel solutions, particularly for regression issues, the goal of this study is to look into the creation of new SVM kernels. To categorize CXR pictures into the three categories of COVID-19, pneumonia, and normal utilizing pre-trained CNN models such as AlexNet, ResNet50, ResNet101, VGG-16, and VGG-19, this study proposes a revolutionary SVM Kernel. The results of the suggested approach show that the updated SVM Kernel may be used as a more effective forecasting tool. ResNet50 offers the greatest accuracy 96.2% and produces the best optimization results in a very short amount of time. [ABSTRACT FROM AUTHOR]

The coronavirus disease (COVID-19) pandemic, originating in Wuhan, China, in 2019, has spread to over 200 countries, posing a significant burden on global health systems and the global economy. The urgent need for faster and more accessible diagnostic methods has become evident, especially in regions with limited resources. Early detection can help reduce the death rate and prevent the spread of disease between people. Reverse transcription polymerase chain reaction (RT-PCR) tests have been widely used as a reliable method for detecting the presence of the virus. However, PCR testing has its drawbacks, such as being time-consuming, expensive, and putting healthcare workers at risk due to close contact with patients. In response to these challenges, medical imaging techniques like X-rays have been utilized for COVID-19 screening due to their speed, safety, and wider availability compared to traditional methods. Integrating artificial intelligence (AI) into image processing has shown promising results in accurately distinguishing between normal and diseased chest radiographs, aiding in early diagnosis and patient management. In this study, we propose a deep learning model to analyze COVID-19 chest X-rays, aiming to address the limitations of PCR testing and enhance the efficiency of diagnosis. One of the major challenges faced in this research is the low amount of available data for training the deep learning model. The lack of comprehensive and well-labeled datasets hinders the model's ability to generalize effectively and may limit its performance. Despite this limitation, we aim to achieve the best possible results using the available data. The proposed method includes a preprocessing step to improve image quality, followed by lung segmentation using UNet to eliminate irrelevant data and improve learning. Data augmentation is applied to address the class imbalance, ensuring the model's generalization capability. The core of the system is a lightweight convolutional neural network (CNN), optimized for small training datasets and low computation cost, making it suitable for resource-limited devices and real-time applications. The classification performance of the proposed model is compared with widely-used CNN models, including DenseNet, ResNet, InceptionV3, VGG16, and VGG19. The results demonstrate the superiority of the lightweight CNN model in terms of computation cost and generalization, showing great promise for its practical implementations. This research contributes to advancing the diagnosis and medical decision support system of COVID-19 through deep learning techniques, particularly in regions where PCR testing is limited or not readily accessible. By providing a faster, reliable, and cost-effective alternative to PCR testing, the proposed method can aid in mitigating the spread of the virus and improve patient outcomes. [ABSTRACT FROM AUTHOR]

Emotion plays a vital role in every living being. Understanding emotion is a very complex task for everyone, but if possible, it will work like a miracle to solve thousands of problems and save many lives. Emotion is reflected not only in the gesture but also in work and in producing an efficient result. Hence, the recognition of emotion using speech has been a topic of interest for many researchers for the last three decades. In our study, we used three features, mel frequency cepstral coefficient (MFCC), spectrogram, and mel-spectrogram, as a one-dimensional input vector to the convolutional neural network (CNN) and deep neural network (DNN) for speech emotion classification. We evaluated the accuracy of SER using the features individually and in combination with the deep learning classifiers CNN and DNN. For both CNN and DNN classifiers, the combination of features performed better than the individual features. The combination of features using the DNN classifier achieved an accuracy of 76.60%, 87.10%, 79.79%, and 100%, and using the CNN classifier achieved classification accuracy of 75%, 84.11%, 78.13%, and 100% for the RAVDESS, EMO-DB, SAVEE, and TESS datasets,respectively. Then we applied a proposed feature and classifier-level fusion method using CNN and DNN to improve emotion classification performance and achieved classification accuracy of 80.42%, 87.48%, and 80.99% on the RAVDESS, EMO-DB, and SAVEE datasets, respectively. The performance of the proposed feature and classifier-level fusion method was compared with the other methods, and it was found that the proposed method performed better than the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Electrocardiogram (ECG) is one among the most common detecting techniques in the analysis and detection of cardiac arrhythmia adopted due to its cost efficiency and simplicity. In a clinical routine, ECG database is collected on daily basis and these databases are reviewed manually. Along with other conventional methods, various approaches using machine learning has been proposed in the past few years. But these would require in-depth knowledge on several parameters and pre-processing techniques in the specific domain. This study is aimed at implementing a more reliable deep learning model that has the capacity to diagnose arrhythmia from a database with 109,446 samples in 5 different categories. In our proposed work, we have used deep learning methodologies for the diagnosis and detection of cardiac arrhythmia automatically. Balancing the biasedness in the waveforms from MIT-BIH arrhythmia database, model is developed. MIT-BIH arrhythmia database with the ECG waveforms promises good accuracy. This automated prediction of the disease using CNN and ResNet-18 architectures are compared in terms of accuracy. CNN has accuracy approximately 97.86% and 98.14% for improved ResNet-18. Also, a comparative analysis is done with the proposed model and already existing techniques. Several limitations and future opportunities are also reviewed. We believe it can be used considerably for cardiac arrhythmia prediction worldwide. Based on the results obtained, ResNet-18 architecture can be used as an efficient procedure, that reduces the burden of training a deep convolutional neural network from start, resulting in a technique that is simple to use. [ABSTRACT FROM AUTHOR]

Mammography is one of the most effective tools radiologists use to detect breast cancer early, as it can detect cancer up to ten years before it manifests. The accuracy of breast cancer segmentation using computer aided design (CAD) systems is challenged by several factors, such as insufficient data for model training and data complexity. For example, mammography images are symmetrical by nature and appear complex to computer vision models, making accurate segmentation difficult. Data augmentation techniques are employed by researchers to help computer vision models preserve the invariance and equivariance features of transformed input images. However, these techniques have been insufficient and incur a computational cost. To improve the data efficiency of the computer vision models, this study proposes a model that combines the functionalities of the Mask Region Convolutional Neural Networks (Mask-RCNN) and Group Convolutional Neural Networks (GCNN) to provide highly accurate breast cancer segmentation. The GCNN preserves the invariant rotation of the transformed mammography images, increases weight sharing and the expressive capacity of the model, and minimizes computational costs. Our proposed model was applied to two mammography datasets, INbreast and MIAS. It achieved competitive results in terms of accuracy, Dice coefficient, Jaccard index, specificity, and sensitivity of 99.01%, 86.63%, 87.76%, 99.24%, and 98.55%, respectively, compared to conventional architectures. [ABSTRACT FROM AUTHOR]

Malware detection has recently emerged as a significant challenge on the Internet of Things (IoT) security domain. Due to the increasing complexity and variety of malware, the demand for better malware detection models has grown. To this end, an advanced intelligent IoT malware detection model is proposed based on deep learning and ensemble learning algorithms, called DEMD-IoT (Deep Ensemble Malware Detection for IoT). The DEMD-IoT includes two main parts: I) a stack of three one-dimensional convolutional neural networks with different structures to learn various patterns of IoT network traffic, and II) a meta-learner on top of three pre-trained base-learners, which integrates the outcomes and yields the final prediction result. To find the best meta-learner, different machine learning algorithms are examined. Finally, the Random Forest algorithm was chosen for the second part of the model. The main advantages of the DEMD-IoT are utilizing an ensemble learning approach to achieve higher performance and applying a hyperparameter optimization algorithm to find the best values for hyperparameters of base learners. Furthermore, to overcome the complexity and high time consumption in the preprocessing phase, 1D-CNNs are applied instead of 2D-CNNs, which are widely used in malware detection. The performance of the DEMD-IoT model is evaluated with a set of experiments on the IoT-23 dataset, which contains IoT network traffic. The findings demonstrate that the proposed ensemble method achieves the best outcome with the highest accuracy 99.9%, compared to state-of-the-art machine learning, deep learning, and ensemble models. [ABSTRACT FROM AUTHOR]