Training Deep Learning Models Research Articles

Enhancing automated strabismus classification with limited data: Data augmentation using StyleGAN2-ADA.

In this study, we propose a generative data augmentation technique to overcome the challenges of severely limited data when designing a deep learning-based automated strabismus diagnosis system. We implement a generative model based on the StyleGAN2-ADA model for system design and assess strabismus classification performance using two classifiers. We evaluate the capability of our proposed method against traditional data augmentation techniques and confirm a substantial enhancement in performance. Furthermore, we conduct experiments to explore the relationship between the diagnosis agreement among ophthalmologists and the generation performance of the generative model. Beyond FID, we validate the generative samples on the classifier to establish their practicality. Through these experiments, we demonstrate that the generative model-based data augmentation improves overall quantitative performance in scenarios of extreme data scarcity and effectively mitigates overfitting issues during deep learning model training.

Anomaly detection and defense techniques in federated learning: a comprehensive review

In recent years, deep learning methods based on a large amount of data have achieved substantial success in numerous fields. However, with increases in regulations for protecting private user data, access to such data has become restricted. To overcome this limitation, federated learning (FL) has been widely utilized for training deep learning models without centralizing data. However, the inaccessibility of FL data and heterogeneity of the client data render difficulty in providing security and protecting the privacy in FL. In addition, the security and privacy anomalies in the corresponding systems significantly hinder the application of FL. Numerous studies have been proposed aiming to maintain the model security and mitigate the leakage of private training data during the FL training phase. Existing surveys categorize FL attacks from a defensive standpoint, but lack the efficiency of pinpointing attack points and implementing timely defenses. In contrast, our survey comprehensively categorizes and summarizes detected anomalies across client, server, and communication perspectives, facilitating easier identification and timely defense measures. Our survey provides an overview of the FL system and briefly introduces the FL security and privacy anomalies. Next, we detail the existing security and privacy anomalies and the methods of detection and defense from the perspectives of the client, server, and communication process. Finally, we address the security and privacy anomalies in non-independent identically distributed cases during FL and summarize the related research progress. This survey aims to provide a systematic and comprehensive review of security and privacy research in FL to help understand the progress and better apply FL in additional scenarios.

Practical Application of Deep Learning in Diagnostic Neuropathology-Reimagining a Histological Asset in the Era of Precision Medicine.

In the past few decades, neuropathology has experienced several paradigm shifts with the introduction of new technologies. Deep learning, a rapidly progressing subfield of machine learning, seems to be the next innovation to alter the diagnostic workflow. In this review, we will explore the recent changes in the field of neuropathology and how this has led to an increased focus on molecular features in diagnosis and prognosis. Then, we will examine the work carried out to train deep learning models for various diagnostic tasks in neuropathology, as well as the machine learning frameworks they used. Focus will be given to both the challenges and successes highlighted therein, as well as what these trends may tell us about future roadblocks in the widespread adoption of this new technology. Finally, we will touch on recent trends in deep learning, as applied to digital pathology more generally, and what this may tell us about the future of deep learning applications in neuropathology.

Deep learning methodologies based on metaheuristics for predictive engine maintenance

Recently, there has been an increase in concerns about the accessibility, security, and reliability of aviation engines. To prevent engine failures which can be quite serious, it is important to take effective measures. The objective is to create a deep learning simulation that can accurately predict an aircraft engine's viability and remaining usefulness using meta-heuristic techniques to improve its performance. These techniques discover the optimal hyper parameters and architecture for the deep learning model. This will help minimize downtime and maintenance costs for the aircraft fleet by handling complex data such as sensor readings and past maintenance records while also adapting to changing conditions over time. Since training deep learning models can be computationally intensive, meta-heuristic methods increase their robustness. The aim is to enhance performance by increasing the accuracy rate and reducing mean squared losses of multiple deep learning methods used for predicting aircraft engine maintenance by hybridizing them with metaheuristic algorithms.

IMNE: Maximizing influence through deep learning-based node embedding in social network

Influence Maximization (IM) is a critical problem in social network analysis and marketing. It involves identifying a subset of nodes in a social network whose activation or influence can lead to the maximal spread of information, ideas, or behaviors within the network. Although many approaches have been developed in the literature to deal with this problem, most of these approaches are ineffective in dealing with large-scale social networks due to free parameters and computational complexity. Embeddings are used to learn low-dimensional representations of nodes in a social network. These embeddings capture the structural and semantic information of nodes and their relationships within the network. By training deep learning models on graph-structured data, node embeddings can capture complex patterns and dependencies in social networks, enabling more effective downstream tasks such as IM. Accordingly, this paper proposes an efficient algorithm to address the IM problem in social networks using deep learning-based Node Embedding (IMNE), which includes shell decomposition, graph/node embedding, and search space reduction as well as the use of local structural features. Our approach combines the power of deep learning for representation learning with the rich structural information present in social networks to address the challenge of IM in complex and dynamic social networks. IMNE uses the Independent Cascade (IC) information diffusion model to determine the labels needed to train the model by calculating the influence of nodes. Experimental results on several real-world networks considering different performance metrics show that IMNE performs better compared to existing baseline and state-of-the-art methods.

Evaluating the Margins of Breast Cancer Tumors by Using Digital Breast Tomosynthesis with Deep Learning: A Preliminary Assessment.

The assessment information of tumor margins is extremely important for the success of the breast cancer surgery and whether the patient undergoes a second operation. However, conducting surgical margin assessments is a time-consuming task that requires pathology-related skills and equipment, and often cannot be provided in a timely manner. To address this challenge, digital breast tomosynthesis technology was utilized to generate detailed cross-sectional images of the breast tissue and integrate deep learning algorithms for image segmentation, achieving an assessment of tumor margins during surgery. this study utilized post-operative tissue samples from 46 patients who underwent breast-conserving treatment, and generated image sets using digital breast tomosynthesis for the training and evaluation of deep learning models. Deep learning algorithms effectively identifying the tumor area. They achieved a Mean Intersection over Union (MIoU) of 0.91, global accuracy of 99%, weighted IoU of 44%, precision of 98%, recall of 83%, F1 score of 89%, and dice coefficient of 93% on the training dataset; for the testing dataset, MIoU was at 83%, global accuracy at 97%, weighted IoU at 38%, precision at 87%, recall rate at 69%, F1 score at 76%, dice coefficient at 86%. The initial evaluation suggests that the deep learning-based image segmentation method is highly accurate in measuring breast tumor margins. This helps provide information related to tumor margins during surgery, and by using different datasets, this research method can also be applied to the surgical margin assessment of various types of tumors.

Semi-supervised image manipulation localization with residual enhancement

Images have become a significant medium for information transmission, while image forensics has garnered widespread attention from researchers. Due to the scarcity of finely annotated images in the field of image manipulation detection and localization, existing methods for locating manipulated images can only utilize a limited amount of data for model training. However, deep learning models typically require a large number of finely annotated images to fully leverage their powerful fitting capabilities. In this paper, we propose a semi-supervised image manipulation localization framework, employing semi-supervised learning to use unannotated images for deep learning model training. To achieve this goal, we first design a residual enhancement module that contains an encoding-decoding structure. The regression target of this branch is the pristine regions in the images to generate the reconstructed images. Next, we perform a residual operation between the reconstructed and original images to roughly locate the anomalous regions and refine the features extracted by the encoder. Secondly, we employ weakly-supervised learning by adding an image-level classification head to the final layer of the encoder. The classification head generates a class active map based on image-level classification results, which further guides the model’s feature augmentation. Finally, we formulate a specialized semi-supervised framework tailored for image manipulation detection, enabling the utilization of a large volume of unannotated data for model training. Extensive experimental results highlight the notable efficacy of the proposed approach, and outperforms state-of-the-art approaches.

ShuffleFL: Addressing Heterogeneity in Multi-Device Federated Learning

Federated Learning (FL) has emerged as a privacy-preserving paradigm for collaborative deep learning model training across distributed data silos. Despite its importance, FL faces challenges such as high latency and less effective global models. In this paper, we propose ShuffleFL, an innovative framework stemming from the hierarchical FL, which introduces a user layer between the FL devices and the FL server. ShuffleFL naturally groups devices based on their affiliations, e.g., belonging to the same user, to ease the strict privacy restriction-"data at the FL devices cannot be shared with others", thereby enabling the exchange of local samples among them. The user layer assumes a multi-faceted role, not just aggregating local updates but also coordinating data shuffling within affiliated devices. We formulate this data shuffling as an optimization problem, detailing our objectives to align local data closely with device computing capabilities and to ensure a more balanced data distribution at the intra-user devices. Through extensive experiments using realistic device profiles and five non-IID datasets, we demonstrate that ShuffleFL can improve inference accuracy by 2.81% to 7.85% and speed up the convergence by 4.11x to 36.56x when reaching the target accuracy.

Conquering class imbalances in deep learning-based segmentation of dental radiographs with different loss functions

ObjectiveThe imbalanced nature of real-world datasets is an ongoing challenge in the field of machine and deep learning. In medicine and in dentistry, most data samples represent patients not affected by pathologies, and on imagery, pathologic image areas are often smaller than healthy ones. Selecting suitable loss functions during deep learning is essential and may help to overcome the resulting imbalance. We assessed six different loss functions for one exemplary task, tooth structure segmentation on bitewing radiographs, for their performance. MethodsSix different loss functions (Focal Loss, Dice Loss, Tversky Loss and hybrid losses of Cross-Entropy and Dice Loss, Focal and Dice Loss, Focal and Generalized Dice Loss) were compared on a tooth structure segmentation task of 1,625 bitewing radiographs. Training was performed using three different model architectures (U-Net, Linknet, DeepLavbV3+) over a 5-fold cross-validation. Tooth structures consisted of the classes (occurrence in% of samples/captures areas measured on pixel level) enamel (100 %/25 %), dentin (100 %/50 %), root canal (100 %/10 %), filling (81 %/8 %) and crown (28 %/5 %). ResultsHybrid loss functions significantly outperformed standalone ones and provided robust results over the different architectures for the classes enamel, dentin, root canal and filling. Specifically, the Dice Focal loss reached high performance to conquer both image level and pixel level class imbalance, respectively. Clinical SignificanceIn dental use cases it is often important to predict minority classes such as pathologies accurately. Using specific loss function may be an effective strategy to overcome data imbalance when training deep learning models.

A novel multi-scale violence and public gathering dataset for crowd behavior classification

Dependable utilization of computer vision applications, such as smart surveillance, requires training deep learning networks on datasets that sufficiently represent the classes of interest. However, the bottleneck in many computer vision applications lies in the limited availability of adequate datasets. One particular application that is of great importance for the safety of cities and crowded areas is smart surveillance. Conventional surveillance methods are reactive and often ineffective in enable real-time action. However, smart surveillance is a key component of smart and proactive security in a smart city. Motivated by a smart city application which aims at the automatic identification of concerning events for alerting law-enforcement and governmental agencies, we craft a large video dataset that focuses on the distinction between small-scale violence, large-scale violence, peaceful gatherings, and natural events. This dataset classifies public events along two axes, the size of the crowd observed and the level of perceived violence in the crowd. We name this newly-built dataset the Multi-Scale Violence and Public Gathering (MSV-PG) dataset. The videos in the dataset go through several pre-processing steps to prepare them to be fed into a deep learning architecture. We conduct several experiments on the MSV-PG datasets using a ResNet3D, a Swin Transformer and an R(2 + 1)D architecture. The results achieved by these models when trained on the MSV-PG dataset, 88.37%, 89.76%, and 89.3%, respectively, indicate that the dataset is well-labeled and is rich enough to train deep learning models for automatic smart surveillance for diverse scenarios.

A rolling bearing fault diagnosis method based on interactive generative feature space oversampling-based autoencoder under imbalanced data

With the rapid development of railroads and the yearly increase in the scale of operation, the safe operation and maintenance of rail trains have become particularly important. Among them, deep learning-based bearing fault diagnosis methods have attracted more and more attention in rail train operation and maintenance. However, rail trains usually operate normally. Collecting complete fault data for deep learning model training is often difficult. Such scenarios with a large difference between the number of normal data and fault data usually affect the performance of fault diagnosis models. Here, an interactive generative feature space oversampling-based autoencoder (IGFSO-AE) is proposed to realize fault sample generation under imbalanced data. First, the original vibration signal is converted into a semantically stable amplitude–frequency signal by fast Fourier transform and input into the autoencoder; second, the order of the hidden layer space features of the autoencoder is randomly exchanged, and the interactive sample generation learning strategy trains the autoencoder; then, interpolation oversampling is used to interpolate samples in the hidden layer space where the Euclidean distance between samples is large, and is input into the decoder, the generated samples are mixed with the original samples to form a new training set, which is used to train the intelligent fault diagnosis model and output the diagnosis results. Finally, the performance of the proposed method is evaluated using the publicly available bearing dataset and the bogie-bearing fault simulation bench in our lab. The experimental results show that IGFSO-AE can generate diverse samples with incremental information and exhibits robustness and superiority in different imbalanced proportions of data.

Enhancing medical image classification with generative AI using latent denoising diffusion probabilistic model and wiener filtering approach

In the evolving landscape of medical diagnostics, a paradigm shift is being catalysed by the advent of generative artificial intelligence. Medical X-ray, CT and MRI images are essential diagnostic tools used by healthcare professionals to assess various musculoskeletal conditions. However, obtaining a sufficient number of medical images for training deep learning models can be challenging due to limited access to labelled data. Hence, the authors propose a novel Latent diffusion process for synthesizing medical images that closely resemble real patient images, aiming to address the challenge of limited access to labelled data in medical diagnostics. Leveraging deep learning and generative modelling techniques, this method synthesizes high-fidelity images that closely mimic real patient scans. By introducing noise and subsequently training the model to denoise, the approach captures intricate patterns inherent in authentic medical images. Among the four datasets utilized, the Diabetes Retinopathy dataset demonstrates superior performance, achieving the highest Mean Structural Similarity Index (MSSIM) score of 0.57 (compared to the dataset baseline of 0.62) and an accuracy of 93.75% when passed through the proposed pipeline. The Cataract dataset, registered a MSSIM score of 0.51 (versus the dataset baseline of 0.53) and an accuracy score of 97.52%, while the Knee OA dataset follows closely with MSSIM and accuracy scores of 0.65 (in contrast to the dataset baseline of 0.63) and 68.66% respectively. The results obtained are then compared with the results generated by the other state of the art models.

Construction of design requirements knowledgebase from unstructured design guidelines using natural language processing

In the manufacturing industry, unstructured documents such as design guidelines, regulatory documents, and failure cases are essential for product development. However, due to the large volume and frequent revisions of these documents, designers often find it difficult to keep up to date with the latest content. This study presents a method for analyzing the characteristics of unstructured design guidelines and automatically constructing a knowledgebase of design requirements from them. A knowledgebase is structured data that a computer can understand, and that can be used to assist designers in the design process. The knowledgebase is constructed using the sections of the document, including design variables and design requirements. The construction process involves pre-processing the documents, extracting information using natural language processing models, and generating a knowledgebase using predefined rules. A requirements knowledgebase was experimentally constructed from a standard document on the general requirements for the design of pressure vessels (American Society of Mechanical Engineers Section VIII Division 1) using the proposed method. In the experiment, the accuracy of information extraction was 86.3 %, and the generation process took 3 min and 50 s. Thus, the proposed method eliminates the need for specialized training of deep learning models and can be applied to various design guideline documents with simple modifications to the design vocabulary and rules. The knowledgebase has applications in design validation, and is expected to enhance the efficiency of the product development process and contribute to reducing the overall development timeline.

Cross-sensor transfer learning for fire smoke scene detection using variable-bands multi-spectral satellite imagery aided by spectral patterns

ABSTRACT This paper addresses the challenge of training deep learning models for fire smoke scene detection from multi-sensor, multi-spectral satellite imagery, where spectral bands vary and training data is scarce, especially for new sensors. We introduce a novel cross-sensor transfer learning approach enhanced by spectral pattern extraction, notably aided by a deep learning (DL) module called Input Amplification (IA). Our contribution includes the creation of two datasets, Landsat_smk2770 and Sentinel2_smk, with varying spectral bands. We employ a model named IA_VIB_SD which incorporates IA and is designed specifically for fire smoke scene detection. We investigated the transferability of the Landsat_smk2770-trained IA_VIB_SD model to the Sentinel2_smk dataset using various transfer learning techniques, and compared the performance of the transferred model with the benchmark IA_VIB_SD model exclusively trained on pure Sentinel2_smk data. Our proposed transfer learning approach resulted in a transferred model with an average accuracy 5% higher than the benchmark model. Notably, Our proposed transfer learning approach outperformed existing transfer learning methods by more than 1% in terms of the accuracy, even when trained on a mere 10% of the Sentinel2_smk dataset. This research advances fire smoke scene detection from multi-sensor, multi-spectral satellite imagery.

LRFE: A NOVEL LOCAL RESPONSE FEATURE ELIMINATION PROCESS FOR IDENTIFICATION OF LUNG CANCER CELLS

One of the main causes of cancer-related mortality across the globe is lung cancer. Early-stage lung cancer frequently exhibits no symptoms, which delays diagnosis until the illness has progressed. Before symptoms manifest, screening and early detection techniques can aid in the early diagnosis and treatment of lung cancer. Many olden research papers have implemented image processing and few latest papers have implemented computer vision techniques to detect lung cancer. Particularly when dealing with minor or subtle anomalies, image processing algorithms may not be able to detect lung cancer lesions with sufficient sensitivity and specificity. It is still difficult to increase the detection algorithms' accuracy and dependability, especially when dealing with early-stage lesions or situations where attributes overlap. It takes a lot of processing power, such as high-performance GPUs and enormous memory capacities, to train deep learning models, particularly large-scale convolutional neural networks (CNNs). In this proposed research, the model pre-processes the images using the ostu and sober filter mechanisms because Otsu's approach adjusts to the features of the input image, including noise, contrast, and lighting fluctuations. It is capable of handling images with varying dynamic ranges and intensity distributions without depending on pre-established threshold settings. When it comes to image noise, the Sobel filter is more resilient than other edge detection methods. It produces clearer edge maps and fewer false detections by determining the gradient magnitude, which amplifies edge information while suppressing noise. The features are extracted using the tuned AlexNet pre-trained model, in AlexNet there is a layer known as “Layer-wise Relevance Propagation”. By giving each pixel or feature in the input image a relevance score, the LRP layer offers fine-grained feature attribution. This makes it possible to analyze in great depth which particular elements or areas of the input image are most important for the network to forecast, which helps to clarify the underlying patterns that the network has learned. At last, the extracted features are further reduced using the enhanced feature elimination method. By iteratively selecting subsets of features based on their importance, RFE helps to identify the most relevant features for the classification task. Integrating RFE with SVM classifiers can lead to improved model performance by focusing on the subset of features that are most discriminative and informative for the classification problem.

Improving Computer-Aided Medical Diagnosis Using Generative Adversarial Networks for Carotid Artery Ultrasound Image Data Augmentation and Classification

Cardiovascular diseases are a leading cause of death globally, making early detection of atherosclerosis critical for prevention. Carotid artery ultrasound imaging is a common diagnostic tool; however, the limited availability of labelled medical images hinders the training of deep learning models. This study examines generative adversarial networks (GANs) for data augmentation and classification of carotid artery Doppler images to improve computer-aided medical diagnosis. Four convolutional neural networks (CNNs) – AlexNet, VGGNet, GoogleNet, and CifarNet – are evaluated for their classification performance on original and extended datasets. AlexNet outperforms the other models, achieving a classification accuracy of 94.18% on the extended dataset. The GAN implementation for data augmentation and overfitting reduction demonstrates the potential of generative models in enhancing the performance of deep learning models in medical image analysis, particularly in the "common artery carotid" class. This research contributes to understanding GANs as a valuable tool for data augmentation and classification in the context of carotid artery ultrasound images.

Deep learning enhanced mixed integer optimization: Learning to reduce model dimensionality

This work introduces a framework to address the computational complexity inherent in Mixed Integer Programming (MIP) models by harnessing the potential of deep learning. By employing deep learning, we construct problem-specific heuristics that identify and exploit common structures across MIP instances. We train deep learning models to estimate complicating binary variables for target MIP problem instances. The resulting reduced MIP models are solved using standard off-the-shelf solvers. We present an algorithm for generating synthetic data enhancing the robustness and generalizability of our models across diverse MIP instances. We compare the effectiveness of (a) feed-forward neural networks (ANN) and (b) convolutional neural networks (CNN). To enhance the framework's performance, we employ Bayesian optimization for hyperparameter tuning, aiming to maximize the occurrence of global optimum solutions. We apply this framework to a flow-based facility location allocation MIP formulation that describes long-term investment planning and medium-term tactical scheduling in a personalized medicine supply chain.

Enhancing ransomware detection using Siamese network

Organizations in the current digital era are exposed to a variety of cybersecurity threats that can often result in financial losses and harm to their reputation. Among these threats, ransomware attacks can cause significant damage. Attackers are constantly improving their techniques to bypass security channels, which makes it challenging to monitor and detect the patterns of attacks. Consequently, there is a growing inclination towards employing state-of-the-art techniques to identify and defend during ransomware attacks. Deep learning is a proven technique that can be employed to learn from large complex patterns. However, large datasets are required in the training of deep learning models which is a challenging task. Few-shot learning (FSL) overcomes this limitation by using less data. In this research work, a Siamese network design is developed by incorporating the architectural principles of AlexNet and features of the VGG configuration. The employed methodology enables us to evaluate the inherent resemblances and disparities in the data. This novel methodology demonstrated exceptional performance, with an average accuracy of 97% when compared to various effects and learning rates. The results of the presented study demonstrate the capacity to greatly enhance cybersecurity by providing a scalable and effective approach for detecting ransomware.

Facial Expression Recognition Using Convolutional Neural Network

Understanding others' intentions through nonverbal cues like facial emotions is crucial in human communication. To design and train Deep Learning Models, this paper describes in detail how Convolutional Neural Network Models are developed using tf. Keras. The aim is to Sort facial photos into one of the seven face detection classifiers, our model is developed in such a manner that it learns hidden nonlinearity from the entered facial images, which is vital for discriminating the form of emotion someone is expressing. The model proposed on the Lenet-5 architecture by Yann LeCun uses the subsampling, feature map, and activation function (ReLu) in between the convolutional layer and fully connected layer for the output soft-max activation function will be used. The FER-2013 dataset, which consists of 35,887 structured 48x48 pixel grayscale images, was used to train the CNN models. The training dataset has 28,709 elements, while the testing dataset has 3,589 elements, and while validation has 3,589 elements. Train and test are the two folder names used to organize the FER dataset. separated even further into distinct files, each holding a different kind of FER dataset class. To mitigate the overfitting of the dropout, batch normalization and the model are employed. Since this is a multiclass classification problem, we are utilizing the Soft-max activation function and the Rectified linear unit for non-linear operation (ReLu). We are training a categorical cross- entropy and matrix for accuracy based on the parameters to assess the constructed CNN model's performance by examining the training epoch history and we have used optimizer Adam (Adaptive Moment Estimation) with the learning rate of 0.0001. We obtained the accuracy of the LeNet-5 model on training at 95.49% and testing at 49.47% [13].

Simulation of Automatically Annotated Visible and Multi-/Hyperspectral Images Using the Helios 3D Plant and Radiative Transfer Modeling Framework.

Deep learning and multimodal remote and proximal sensing are widely used for analyzing plant and crop traits, but many of these deep learning models are supervised and necessitate reference datasets with image annotations. Acquiring these datasets often demands experiments that are both labor-intensive and time-consuming. Furthermore, extracting traits from remote sensing data beyond simple geometric features remains a challenge. To address these challenges, we proposed a radiative transfer modeling framework based on the Helios 3-dimensional (3D) plant modeling software designed for plant remote and proximal sensing image simulation. The framework has the capability to simulate RGB, multi-/hyperspectral, thermal, and depth cameras, and produce associated plant images with fully resolved reference labels such as plant physical traits, leaf chemical concentrations, and leaf physiological traits. Helios offers a simulated environment that enables generation of 3D geometric models of plants and soil with random variation, and specification or simulation of their properties and function. This approach differs from traditional computer graphics rendering by explicitly modeling radiation transfer physics, which provides a critical link to underlying plant biophysical processes. Results indicate that the framework is capable of generating high-quality, labeled synthetic plant images under given lighting scenarios, which can lessen or remove the need for manually collected and annotated data. Two example applications are presented that demonstrate the feasibility of using the model to enable unsupervised learning by training deep learning models exclusively with simulated images and performing prediction tasks using real images.