A Deep Ensemble Learning Approach Based on a Vision Transformer and Neural Network for Multi-Label Image Classification

目的 基于计算机的胸腔X线影像疾病检测和分类目前存在误诊率高，准确率低的问题。本文在视觉Transformer（vision Transformer，ViT）预训练模型的基础上，通过迁移学习方法，实现胸腔X线影像辅助诊断，提高诊断准确率和效率。方法 选用带有卷积神经网络（convolutional neural network，CNN）的ViT模型，其在超大规模自然图像数据集中进行了预训练；通过微调模型结构，使用预训练的ViT模型参数初始化主干网络，并迁移至胸腔X线影像数据集中再次训练，实现疾病多标签分类。结果 在IU X-Ray数据集中对ViT迁移学习前、后模型平均AUC（area under ROC curve）得分进行对比分析实验。结果表明，预训练ViT模型平均AUC得分为0.774，与不使用迁移学习相比提升了0.208。并针对模型结构和数据预处理进行了消融实验，对ViT中的注意力机制进行可视化，进一步验证了模型有效性。最后使用Chest X-Ray14和CheXpert数据集训练微调后的ViT模型，平均AUC得分为0.839和0.806，与对比方法相比分别有0.014～0.031的提升。结论 与其他方法相比，ViT模型胸腔X线影像的多标签分类精确度更高，且迁移学习可以在降低训练成本的同时提升ViT模型的分类性能和泛化性。消融实验与模型可视化表明，包含CNN结构的ViT模型能重点关注有意义的区域，高效获取胸腔X线影像的视觉特征。;Objective The chest X-ray-relevant screening and diagnostic method is essential for radiology nowadays. Most of chest X-ray images interpretation is still restricted by clinical experience and challenged for misdiagnose and missed diagnoses. To detect and identify one or more potential diseases in images automatically，it is beneficial for improving diagnostic efficiency and accuracy using computer-based technique. Compared to natural images，multiple lesions are challenged to be detected and distinguished accurately in a single image because abnormal areas have a small proportion and complex representations in chest X-ray images. Current convolutional neural network（CNN）based deep learning models have been widely used in the context of medical imaging. The structure of the CNN convolution kernel has sensitive to local detail information，and it is possible to extract richer image features. However，the convolution kernel cannot be used to get global information，and the features-extracted are restricted of redundant information like its relevance of background， muscles，and bones. The model’s performance in multi-label classification tasks are affected to a certain extent. At present，the vision Transformer（ViT）model has achieved its priorities in computer vision-related tasks. The ViT can be used to capture information simultaneously and effectively for multiple regions of the entire image. However，it is required to use large-scale dataset training to achieve good performance. Due to some factors like patient privacy and manual annotate costs，the size of the chest X-ray image data set has been limited. To reduce the model's dependence on data scale and improve the performance of multi-label classification，we develop the CNN-based ViT pre-training model in terms of the transfer learning method for diagnosis-assisted of chest X-ray image and multi-label classification. Method The CNN-based ViT model is pre-trained on a huge scale ground truth dataset，and it is used to obtain the initial parameters of the model. The model structure is fine-tuned according to the features of chest X-ray dataset. A 1×1 convolution layer is used to convert the chest X-ray images channels between 1 to 3. The number of output nodes of the linear layer in the classifier is balanced from 1 000 to the number of chest X-ray classification labels，and the Sigmoid is used as an activation function. The parameters of the backbone network are initialized in terms of the pre-trained ViT model parameters，and it is trained in the chest X-ray dataset after that to complete multi-label classification. The experiment is configured of Python3. 7 and PyTorch1. 8 to construct the model and RTX3090 GPU for training. Stochastic gradient descent（SGD）optimizer，binary cross-entropy（BCE）loss function，an initial learning rate of 1E-3，the cosine annealing learning rate decay are used. For training，each image is scaled to a size of 512×512 pixels，and a 224×224 pixels area and it is then cropped in random as the model input，and data augmentation is performed randomly by some of the flipping，perspective transformation， shearing，translation，zooming，and changing brightness. For testing，the chest X-ray image is scaled to 256×256 pixels and center crop a 224×224 area to input the trained model. Result The experiment is performed on the IU X-Ray，which is a small-scale chest X-ray dataset. This model is evaluated in quantitative using the average of area under ROC curve （AUC）scores across all classification labels. The results show that the average AUC score of the pre-trained ViT model is 0. 774. The accuracy and training efficiency of the non-pre-trained ViT model is dropped significantly. The average AUC score is reached to 0. 566 only，which is 0. 208 lower. In addition，the attention mechanism heat map is generated based on the ViT model，which can strengthen the interpretability of the model. A series of ablation experiments are carried out for data augmentation，model structure，and batch size design. The fine-tuned ViT model is trained on the Chest-Ray14 and CheXpert dataset as well. The average AUC score is reached to 0. 839 and 0. 806，which is optimized by 0. 014 and 0. 031. Conclusion A pre-trained ViT model is used for the multi-label classification of chest X-ray images via transfer learning. The experimental results illustrate that the ViT has its stronger multi-label classification performance in chest Xray images，and its attention mechanism is beneficial for lesions precision-focused like the interior of the chest cavity and the heart. Transfer learning is potential to improve the classification performance and model generalization of the ViT in small-scale datasets，and the training cost is reduced greatly. Ablation experiments demonstrate that the incorporated model of CNN and Transformer has its priority beyond single-structure model. Data enhancement and the batch size cutting can improve the performance of the model，but smaller scale of batch is still interlinked to longer training span. To improve the model's ability，we predict that future research direction can be focused on the extraction for complex disease and highlevel semantic information，such as their small lesions，disease location，and severity.

Multi-label image classification focuses on predicting a set of object labels presented in an image, which is a practical yet challenging task since labels are normally co-occurred in an image while mutual interactions among labels and the correspondence between a given image and the corresponding labels are rarely considered in the existing methods. To address above challenges, we propose a Multiple Semantic Embedding model with Graph Convolutional Networks (MSEGCN) for multi-label image classification by capturing important dependencies related labels. Specifically, the proposed MSEGCN leverages graph structure to guide label co-occurrence propagation among different categories to obtain appropriate label representations. Then, by formulating multi-label classification problem as a label ranking problem with the aid of end-to-end convolutional neural network framework, we focus on learning a transform matrix to seek the image-label relevance relations in embedding space. Furthermore, an adaptive weighting strategy is introduced to effectively improve the classification performance. Experimental studies across a wide range of benchmark datasets show that our method achieves highly competitive performances against other state-of-the-art approaches.

Convolutional Neural Networks (CNNs) demonstrate state-of-the-art performance in large-scale multi-class image classification tasks. CNNs consist of convolution layers that progressively construct features and a classification layer. Typically, a softmax function is used in the classification layer to learn joint probabilities for the classes, which are subsequently used for class prediction. We refer to such an approach as the joint approach to multi-class classification. There exists another approach in the literature which determines the multi-class prediction outcome through a sequence of binary decisions, and is christened the class binarization approach. A popular type of class binarization is error-correcting output codes. In this paper, we propose to incorporate error-correcting output codes into convolutional neural networks by inserting a latent binarization layer in a CNN classification layer. This approach encapsulates both encoding and decoding steps of error-correcting output codes into a single CNN architecture that is capable of discovering an optimal coding matrix during training. The latent binarization layer is motivated by the family of latent-trait and latent-class models used in behavioral research. We call the proposed convolutional neural networks with Latent Binarization as LB-CNNs, and develop algorithms combining expectation maximization algorithms and back-propagation to train LB-CNNs. The proposed models and algorithms are applied to several image recognition tasks, producing excellent results. Furthermore, our model enhances the interpretability of the decision process of standard convolutional neural networks.

Iterative fusion convolutional neural networks for classification of optical coherence tomography images

Objectives: The paper presents application of convolution neural network and artificial neural network for image classification problem for clothing dataset along with their performance comparison against different optimizers. The major objective of this paper is to perform image classification on fashion-mnist clothing dataset images. Methods: The methods used here are, the traditional ANN and CNN. Here image classification is performed on Fashion-mnist, clothing dataset using CNN and ANN with different optimizers. The performance of the working of ANN and CNN in classifying images from fashion-mnist dataset is compared against different optimizers namely stochastic gradient Descent, Adagrad, RMS prop and Adam optimizer. Findings: The study found that CNN worked better than ANN yielding training accuracy of 95%, 93% and testing accuracy of 91%, 89% when used with Adam and RmsProp respectively. Novelty: The novelty of this work is to present a comparative study of image classification using CNN, ANN using different optimizers, since not many studies or research articles showed the performance comparison of traditional and convolution neural networks in image classification along with different optimizers. Since the real-world scenarios of today require enormous data to be processed, CNN can fit well to diversify applications since they highly reduce the number of parameters to be trained that speeds up the training process. Moreover, to be specific on image classification problems they require the best and most prominent features to be detected and uncovered; this can be achieved using CNN since it has the concept of convolution using filters at its Core. Hence, CNN is highly recommended for such image classification applications than the traditional artificial-neural-networks because of the aforementioned reasons. Keywords: ANN (Artificial NeuralNetworks); CNN (ConvolutionNeural Networks); Optimization; SGD (Stochastic Gradient Descent); RmsProp (Root mean Square propagation); Adam (Adaptive moment estimation); AdaGrad (Adaptive Gradient)

Deep learning can discover intricate patterns hidden in big data, and has much better scalability than traditional machine learning when the volume of data increases dramatically. Thus, deep learning has gained many successes in various domains and applications such as image classification, text classification, and machine translation. In this paper, we use deep learning to classify geographical features (e.g., mountains, rivers, landmarks, and cities) from text, using geolocated Wikipedia entries as the case study application. We employ one of the most commonly used deep learning architectures, convolutional neural networks (CNNs) and its integration with active learning (creating what we call active CNNs), to train the geographical feature classifiers on the Wikipedia text data set obtained from GeoNames (which provides the feature type for each geolocated entity). We evaluate the performance of CNNs and active CNNs with multiple metrics (i.e., accuracy, F1 score, and confusion matrix). Our experiment results demonstrated that CNNs and active CNNs can effectively classify geo-referenced text entities into predefined geographical features. In addition, our experiment results show that active CNNs outperform CNNs for hard to distinguish classes. In our experiment, we also compared results for hierarchical multi-class classification and flat multiclass classification, and the results show that hierarchical multiclass classification significantly outperforms flat multi-class classification for the data set we used.

The classification of logos is a particular case within computer vision since they have their own characteristics. Logos can contain only text, iconic images or a combination of both, and they usually include figurative symbols designed by experts that vary substantially besides they may share the same semantics. This work presents a method for multi-label classification and retrieval of logo images. For this, Convolutional Neural Networks (CNN) are trained to classify logos from the European Union TradeMark (EUTM) dataset according to their colors, shapes, sectors and figurative designs. An auto-encoder is also trained to learn representations of the input images. Once trained, the neural codes from the last convolutional layers in the CNN and the central layer of the auto-encoder can be used to perform similarity search through kNN, allowing us to obtain the most similar logos based on their color, shape, sector, figurative elements, overall features, or a weighted combination of them provided by the user. To the best of our knowledge, this is the first multi-label classification method for logos, and the only one that allows retrieving a ranking of images with these criteria provided by the user.

Recently, convolutional neural networks (CNNs) have been successfully utilized in polarimetric synthetic aperture radar (PolSAR) image classification and obtained promising results. However, most CNN-based classification methods require a large number of labeled samples and it is difficult to obtain sufficient labeled samples. For this reason, an ensemble dual-branch CNN (EDb-CNN) is proposed for PolSAR image classification with small samples. First, to solve the problem of the small sample in PolSAR image classification, a new data enhancement method based on the superpixel algorithm is proposed to expand the number of labeled samples. Second, to obtain different scales of features from PolSAR images, a Db-CNN model is proposed. This model contains two parallel CNN structures. One CNN branch is used to extract the polarization features from the complex coherency matrix. The other branch is utilized to extract the spatial features based on weighted spatial neighborhood. On the top of these two branches, a feature fusion model is adopted to combine these two deep features, and a weighted loss function is employed to improve the learning procedure. Then, the ensemble learning algorithm is used for each CNN branch and Db-CNN network to obtain the better classification results. Finally, a postprocess algorithm based on the superpixel algorithm is proposed to improve the consistency of classification results. Experiments on two PolSAR datasets show that the proposed method achieves a much better performance than other classification methods, especially when only a few labeled samples are available.

While deep convolutional neural networks (CNNs) have shown a great success in single-label image classification, it is important to note that real world images generally contain multiple labels, which could correspond to different objects, scenes, actions and attributes in an image. Traditional approaches to multi-label image classification learn independent classifiers for each category and employ ranking or thresholding on the classification results. These techniques, although working well, fail to explicitly exploit the label dependencies in an image. In this paper, we utilize recurrent neural networks (RNNs) to address this problem. Combined with CNNs, the proposed CNN-RNN framework learns a joint image-label embedding to characterize the semantic label dependency as well as the image-label relevance, and it can be trained end-to-end from scratch to integrate both information in a unified framework. Experimental results on public benchmark datasets demonstrate that the proposed architecture achieves better performance than the state-of-the-art multi-label classification model

In recent years, the evolution of convolutional neural networks (CNNs) has paved the way for the development of deep-learning-based visual inspection techniques to identify damages in concrete bridges. However, the coexistence of diverse damage attributes within a single image complicates the comprehensive interpretation of these images during practical bridge inspections, posing a persistent challenge for existing single classification tasks. To address this issue, we introduce an innovative algorithm that synergizes graph convolutional network (GCN) and multiaxis vision transformer (MaxViT). Our approach leverages MaxViT to extract image features and subsequently models the co-occurrence correlation of labels. In addition, we use a graph CNN to extract features that capture the correlation between labels. Ultimately, we integrate these image features to enable multilabel classification of damage images. To evaluate the efficacy of our method, we curated a multilabel damage classification dataset for concrete bridges, consisting of 2098 samples encompassing the four prevalent types of spall, rebar, corrosion, and crack observed during bridge inspections. Experimental results showcase that our proposed method achieves a mean average precision (mAP) of 99.37% when applied to real bridge inspection images. In terms of general evaluation metrics, our method attains an accuracy of 98.29%, precision of 96.28%, recall of 97.08%, and F1-score of 96.68%. These findings underscore the superior performance of our method compared with current advanced multilabel classification networks, demonstrating higher accuracy. An interpretable analysis of the classification feature maps generated by our model, using the Grad-CAM algorithm, reveals that the model effectively directs attention to specific regions where damage is located. Consequently, our method holds promising engineering applications and has the potential to significantly enhance the accuracy of support provided for bridge inspection and maintenance endeavors.

With advancements in software development and artificial intelligence, defect prediction has gradually become an essential component of the software development lifecycle. Historically, defect prediction has been considered a multiclass classification problem because defect classes are mutually exclusive. However, software defects can belong to multiple categories simultaneously, making multilabel classification a more appropriate approach. A defect report typically contains the title, body, comments from developers and testers, and code snippets. We used these data items and performed data wrangling on these data to create a holistic summary of the defect report, which contains all the vital information that is useful for defect predictions. In this study, we investigated the multilabel dimension of the defect and performed multilabel classification using machine learning and deep learning techniques while considering the class imbalance and correlations between the labels. In the traditional classification methods, we used three classifiers: Multinomial Naive Bayes, Logistic Regression, and Random Forest. Multilayer Perceptron (MLP) and a Convolutional Neural Network (CNN) with Classifier Chains are used in deep learning. To check the dataset quality, appropriate feature selection, and data dimensionality reduction, we used the chi-square test. To handle the class imbalance, we used Non-Negative Least Squares (NNLS). Our experimental investigations showed significant improvements in the model performance across machine learning and deep learning once the dataset was balanced before training the models. Visual plots of evaluation metrics, such as Hamming loss, Recall, Precision, and F1-score, clearly demonstrated the analysis outcome.

The field of histopathological image analysis has evolved significantly with the advent of digital pathology, leading to the development of automated models capable of classifying tissues and structures within diverse pathological images. Artificial intelligence algorithms, such as convolutional neural networks, have shown remarkable capabilities in pathology image analysis tasks, including tumor identification, metastasis detection, and patient prognosis assessment. However, traditional manual analysis methods have generally shown low accuracy in diagnosing colorectal cancer using histopathological images. This study investigates the use of AI in image classification and image analytics using histopathological images using the histogram of oriented gradients method. The study develops an AI-based architecture for image classification using histopathological images, aiming to achieve high performance with less complexity through specific parameters and layers. In this study, we investigate the complicated state of histopathological image classification, explicitly focusing on categorizing nine distinct tissue types. Our research used open-source multi-centered image datasets that included records of 100.000 non-overlapping images from 86 patients for training and 7180 non-overlapping images from 50 patients for testing. The study compares two distinct approaches, training artificial intelligence-based algorithms and manual machine learning models, to automate tissue classification. This research comprises two primary classification tasks: binary classification, distinguishing between normal and tumor tissues, and multi-classification, encompassing nine tissue types, including adipose, background, debris, stroma, lymphocytes, mucus, smooth muscle, normal colon mucosa, and tumor. Our findings show that artificial intelligence-based systems can achieve 0.91 and 0.97 accuracy in binary and multi-class classifications. In comparison, the histogram of directed gradient features and the Random Forest classifier achieved accuracy rates of 0.75 and 0.44 in binary and multi-class classifications, respectively. Our artificial intelligence-based methods are generalizable, allowing them to be integrated into histopathology diagnostics procedures and improve diagnostic accuracy and efficiency. The CNN model outperforms existing machine learning techniques, demonstrating its potential to improve the precision and effectiveness of histopathology image analysis. This research emphasizes the importance of maintaining data consistency and applying normalization methods during the data preparation stage for analysis. It particularly highlights the potential of artificial intelligence to assess histopathological images.

Designing and developing automated systems to detect and grade Diabetic Retinopathy (DR) is one of the recent research areas in the world of medical image applications since it is considered one of the main causes of total blindness for people who have diabetes in the mid-age. In this paper, a complete pipeline for retinal fundus images processing and analysis has been described, implemented and evaluated. This pipeline has three main stages: (i) image pre-processing, (ii) features extraction and (iii) classification. In the first stage, the image has been pre-processed using different transformations to standardize the images and to enhance the images quality. It has been proven that Gaussian filtering is quite effective in this context to enhance the images contrast. In the second and third stage, the convolution neural network (CNN), one of the best neural network architecture for image analysis applications, has been used. The concept of transfer learning and fine tuning have been advocated in this paper and applied for ResNet18 using the publicly available Kaggle dataset. The problem of DR diagnosis has been handled as a multi-class classification problem where there are five levels of the disease severity (0 – No DR, 1 – Mild, 2 – Moderate, 3 – Severe, 4 – Proliferative DR). The final model has achieved accuracy of 70 %, recall of 50% and specificity of 88% outperforming other models built from scratch with less training time and proving the efficiency of transfer learning in this context. The training process has considered the problem of imbalanced dataset using two different ways and it has been discovered that using imbalanced dataset sampler is a very efficient solution. The final model developed in this research could be used as the main unit for a computer aided system to be hosted online for DR detection and diagnosis.

Supervised learning-based methods for source localization, being data driven, can be adapted to different acoustic conditions via training and have been shown to be robust to adverse acoustic environments. In this paper, a convolutional neural network (CNN) based supervised learning method for estimating the direction of arrival (DOA) of multiple speakers is proposed. Multi-speaker DOA estimation is formulated as a multi-class multi-label classification problem, where the assignment of each DOA label to the input feature is treated as a separate binary classification problem. The phase component of the short-time Fourier transform (STFT) coefficients of the received microphone signals are directly fed into the CNN, and the features for DOA estimation are learnt during training. Utilizing the assumption of disjoint speaker activity in the STFT domain, a novel method is proposed to train the CNN with synthesized noise signals. Through experimental evaluation with both simulated and measured acoustic impulse responses, the ability of the proposed DOA estimation approach to adapt to unseen acoustic conditions and its robustness to unseen noise type is demonstrated. Through additional empirical investigation, it is also shown that with an array of M microphone our proposed framework yields the best localization performance with M-1 convolution layers. The ability of the proposed method to accurately localize speakers in a dynamic acoustic scenario with varying number of sources is also shown.

Bond rating is regarded as an important event for measuring financial risk of companies and for determining the investment returns of investors. As a result, it has been a popular research topic for researchers to predict companies' credit ratings by applying statistical and machine learning techniques. The statistical techniques, including multiple regression, multiple discriminant analysis (MDA), logistic models (LOGIT), and probit analysis, have been traditionally used in bond rating. However, one major drawback is that it should be based on strict assumptions. Such strict assumptions include linearity, normality, independence among predictor variables and pre-existing functional forms relating the criterion variablesand the predictor variables. Those strict assumptions of traditional statistics have limited their application to the real world. Machine learning techniques also used in bond rating prediction models include decision trees (DT), neural networks (NN), and Support Vector Machine (SVM). Especially, SVM is recognized as a new and promising classification and regression analysis method. SVM learns a separating hyperplane that can maximize the margin between two categories. SVM is simple enough to be analyzed mathematical, and leads to high performance in practical applications. SVM implements the structuralrisk minimization principle and searches to minimize an upper bound of the generalization error. In addition, the solution of SVM may be a global optimum and thus, overfitting is unlikely to occur with SVM. In addition, SVM does not require too many data sample for training since it builds prediction models by only using some representative sample near the boundaries called support vectors. A number of experimental researches have indicated that SVM has been successfully applied in a variety of pattern recognition fields. However, there are three major drawbacks that can be potential causes for degrading SVM's performance. First, SVM is originally proposed for solving binary-class classification problems. Methods for combining SVMs for multi-class classification such as One-Against-One, One-Against-All have been proposed, but they do not improve the performance in multi-class classification problem as much as SVM for binary-class classification. Second, approximation algorithms (e.g. decomposition methods, sequential minimal optimization algorithm) could be used for effective multi-class computation to reduce computation time, but it could deteriorate classification performance. Third, the difficulty in multi-class prediction problems is in data imbalance problem that can occur when the number of instances in one class greatly outnumbers the number of instances in the other class. Such data sets often cause a default classifier to be built due to skewed boundary and thus the reduction in the classification accuracy of such a classifier. SVM ensemble learning is one of machine learning methods to cope with the above drawbacks. Ensemble learning is a method for improving the performance of classification and prediction algorithms. AdaBoost is one of the widely used ensemble learning techniques. It constructs a composite classifier by sequentially training classifiers while increasing weight on the misclassified observations through iterations. The observations that are incorrectly predicted by previous classifiers are chosen more often than examples that are correctly predicted. Thus Boosting attempts to produce new classifiers that are better able to predict examples for which the current ensemble's performance is poor. In this way, it can reinforce the training of the misclassified observations of the minority class. This paper proposes a multiclass Geometric Mean-based Boosting (MGM-Boost) to resolve multiclass prediction problem. Since MGM-Boost introduces the notion of geometric mean into AdaBoost, it can perform learning process considering the geometric mean-based accuracy and errors of multiclass. This study applies MGM-Boost to the real-world bond rating case for Korean companies to examine the feasibility of MGM-Boost. 10-fold cross validations for threetimes with different random seeds are performed in order to ensure that the comparison among three different classifiers does not happen by chance. For each of 10-fold cross validation, the entire data set is first partitioned into tenequal-sized sets, and then each set is in turn used as the test set while the classifier trains on the other nine sets. That is, cross-validated folds have been tested independently of each algorithm. Through these steps, we have obtained the results for classifiers on each of the 30 experiments. In the comparison of arithmetic mean-based prediction accuracy between individual classifiers, MGM-Boost (52.95%) shows higher prediction accuracy than both AdaBoost (51.69%) and SVM (49.47%). MGM-Boost (28.12%) also shows the higher prediction accuracy than AdaBoost (24.65%) and SVM (15.42%)in terms of geometric mean-based prediction accuracy. T-test is used to examine whether the performance of each classifiers for 30 folds is significantly different. The results indicate that performance of MGM-Boost is significantly different from AdaBoost and SVM classifiers at 1% level. These results mean that MGM-Boost can provide robust and stable solutions to multi-classproblems such as bond rating.