Robust Classification Research Articles

A Wushu Leg Gesture Recognition Algorithm Based on Random Forest and Bone Feature Extraction (RF-SFE)

Addressing the challenges of complex movement trajectories and rapid action changes in martial arts performances, this study introduces a novel posture recognition algorithm based on Random Forest and Skeletal Feature Extraction (RF-SFE) for martial arts leg movements. Unlike traditional posture recognition methods that struggle with accuracy, RF-SFE aims to provide intelligent analysis of training postures to assist practitioners in efficient training. The algorithm initially employs advanced skeletal feature extraction techniques to identify and articulate the relative positions and movements specific to martial arts. These extracted spatial features enhance the flexibility in modeling the unique dynamics of martial arts. Subsequently, Random Forest classification is utilized to categorize different leg movements, leveraging its strength in handling high-dimensional data and providing robust classification. Comparative experiments on diverse martial arts posture datasets demonstrate a significant improvement in recognition rates over baseline methods. This validates the effectiveness of the RF-SFE method in recognizing martial arts postures, offering scientific guidance for practitioners’ training regimens.

A Comprehensive Approach to Fault Classification of Helicopter Engines with Adaboost Ensemble Model

This work is based on the PHM North America 2024 Conference Data Challenge’s datasets of Helicopter turbine engine performance measurements. These datasets were large and moderately imbalanced. For dealing with these challenges, we demonstrate significant tools covering feature engineering, augmentation and selection, model exploration, visualizations, model explainability and confidence margin estimation. This work was performed in its entirety using MATLAB. All these tools will be generally applicable to data-driven modeling and prediction of health to real life applications. Initially, we explored the 742k observations in the training set, noting a 60-40 split between healthy and faulty labels, and identified two major operational clusters within the data. We enhanced the dataset by removing duplicates and engineered new features based on domain knowledge, expanding the feature set to 242 dimensions. However for the torque margin estimation, we trained a regression model on a limited subset (18 features), which includes engineered features using domain knowledge, quadratic terms and linear interaction between all the terms. For the final submission, we utilized a stepwise linear regression model to optimize feature selection. This approach achieved a perfect regression score on test data, validated by a consistent torque margin residual range of +/- 0.5%. The model's RMSE and MAE metrics were optimal for employing a normal distribution probability density function. For , we reduced the feature set to 58 using dimensionality reduction techniques and balanced the data with upsampling and down-weighing the minority class. We employed ASHA (Asynchronous Successive Halving Algorithm) in conjunction with AutoML to efficiently determine the most suitable model family, significantly saving compute time. Subsequently, we trained ensemble models, including bagged tree and AdaBoost (Adaptive Boosting), which minimized false negatives and positives, achieving robust classification performance. This was particularly critical given the high penalty for false negatives in the data challenge. The MathWorks team score on Testing Data was 0.9686 at the close of competition. This was further improved to 0.9867. Our approach demonstrates the effectiveness of combining strategic data processing, feature engineering, and model selection to enhance predictive accuracy in complex operational datasets.

Improving CNN Fish Detection and Classification with Tracking

The regular and consistent monitoring of marine ecosystems and fish communities is becoming more and more crucial due to increasing human pressures. To this end, underwater camera technology has become a major tool to collect an important amount of marine data. As the size of the data collected outgrew the ability to process it, new means of automatic processing have been explored. Convolutional neural networks (CNNs) have been the most popular method for automatic underwater video analysis for the last few years. However, such algorithms are rather image-based and do not exploit the potential of video data. In this paper, we propose a method of coupling video tracking and CNN image analysis to perform a robust and accurate fish classification on deep sea videos and improve automatic classification accuracy. Our method fused CNNs and tracking methods, allowing us to detect 12% more individuals compared to CNN alone.

Optimizing classification models for medical image diagnosis: a comparative analysis on multi-class datasets

The surge in machine learning (ML) and artificial intelligence has revolutionized medical diagnosis, utilizing data from chest ct-scans, COVID-19, lung cancer, brain tumor, and alzheimer parkinson diseases. However, the intricate nature of medical data necessitates robust classification models. This study compares support vector machine (SVM), naïve Bayes, k-nearest neighbors (K-NN), artificial neural networks (ANN), and stochastic gradient descent on multi-class medical datasets, employing data collection, Canny image segmentation, hu moment feature extraction, and oversampling/under-sampling for data balancing. Classification algorithms are assessed via 5-fold cross-validation for accuracy, precision, recall, and F-measure. Results indicate variable model performance depending on datasets and sampling strategies. SVM, K-NN, ANN, and SGD demonstrate superior performance on specific datasets, achieving accuracies between 0.49 to 0.57. Conversely, naïve Bayes exhibits limitations, achieving precision levels of 0.46 to 0.47 on certain datasets. The efficacy of oversampling and under-sampling techniques in improving classification accuracy varies inconsistently. These findings aid medical practitioners and researchers in selecting suitable models for diagnostic applications.

Detection of Subarachnoid Hemorrhage Using CNN with Dynamic Factor and Wandering Strategy-Based Feature Selection

Objectives: Subarachnoid Hemorrhage (SAH) is a serious neurological emergency case with a higher mortality rate. An automatic SAH detection is needed to expedite and improve identification, aiding timely and efficient treatment pathways. The existence of noisy and dissimilar anatomical structures in NCCT images, limited availability of labeled SAH data, and ineffective training causes the issues of irrelevant features, overfitting, and vanishing gradient issues that make SAH detection a challenging task. Methods: In this work, the water waves dynamic factor and wandering strategy-based Sand Cat Swarm Optimization, namely DWSCSO, are proposed to ensure optimum feature selection while a Parametric Rectified Linear Unit with a Stacked Convolutional Neural Network, referred to as PRSCNN, is developed for classifying grades of SAH. The DWSCSO and PRSCNN surpass current practices in SAH detection by improving feature selection and classification accuracy. DWSCSO is proposed to ensure optimum feature selection, avoiding local optima issues with higher exploration capacity and avoiding the issue of overfitting in classification. Firstly, in this work, a modified region-growing method was employed on the patient Non-Contrast Computed Tomography (NCCT) images to segment the regions affected by SAH. From the segmented regions, the wide range of patterns and irregularities, fine-grained textures and details, and complex and abstract features were extracted from pre-trained models like GoogleNet, Visual Geometry Group (VGG)-16, and ResNet50. Next, the PRSCNN was developed for classifying grades of SAH which helped to avoid the vanishing gradient issue. Results: The DWSCSO-PRSCNN obtained a maximum accuracy of 99.48%, which is significant compared with other models. The DWSCSO-PRSCNN provides an improved accuracy of 99.62% in CT dataset compared with the DL-ICH and GoogLeNet + (GLCM and LBP), ResNet-50 + (GLCM and LBP), and AlexNet + (GLCM and LBP), which confirms that DWSCSO-PRSCNN effectively reduces false positives and false negatives. Conclusions: the complexity of DWSCSO-PRSCNN was acceptable in this research, for while simpler approaches appeared preferable, they failed to address problems like overfitting and vanishing gradients. Accordingly, the DWSCSO for optimized feature selection and PRSCNN for robust classification were essential for handling these challenges and enhancing the detection in different clinical settings.

MosquitoSong+: A noise-robust deep learning model for mosquito classification from wingbeat sounds.

In order to assess risk of mosquito-vector borne disease and to effectively target and monitor vector control efforts, accurate information about mosquito vector population densities is needed. The traditional and still most common approach to this involves the use of traps along with manual counting and classification of mosquito species, but the costly and labor-intensive nature of this approach limits its widespread use. Numerous previous studies have sought to address this problem by developing machine learning models to automatically identify species and sex of mosquitoes based on their wingbeat sounds. Yet little work has addressed the issue of robust classification in the presence of environmental background noise, which is essential to making the approach practical. In this paper, we propose a new deep learning model, MosquitoSong+, to identify the species and sex of mosquitoes from raw wingbeat sounds so that it is robust to the environmental noise and the relative volume of the mosquito's flight tone. The proposed model extends the existing 1D-CNN model by adjusting its architecture and introducing two data augmentation techniques during model training: noise augmentation and wingbeat volume variation. Experiments show that the new model has very good generalizability, with species classification accuracy above 80% on several wingbeat datasets with various background noise. It also has an accuracy of 93.3% for species and sex classification on wingbeat sounds overlaid with various background noises. These results suggest that the proposed approach may be a practical means to develop classification models that can perform well in the field.

Classification of Non-Seismic Tsunami Early Warning Level Using Decision Tree Algorithm

Background: Tsunami caused by volcanic collapse are categorized as non-seismic uncommon events, unlike tsunamis caused by earthquakes, which are common events. The traditional tsunami early warning based on the seismic sensor (e.g. earthquake detectors) may not be applicable to volcanic tsunamis because they do not generate seismic waves. Consequently, these tsunamis cannot be detected in advance, and warnings cannot be issued. New methods should be explored to address these non-seismic tsunamis caused by volcanic collapse. Objective: This study explored the potential of machine learning algorithms in supporting early warning level issuing for non-seismic tsunamis, specifically volcanic tsunamis. The Anak Krakatau volcano event in Indonesia was used as a case study. Methods: This study generated a database of 160 collapse scenarios using numerical simulation as input sequences. A classification model was constructed by defining the worst tsunami elevation and its arrival time at the coast. The database was supervised by labeling the warning levels as targets. Subsequently, a decision tree algorithm was employed to classify the warning levels. Results: The results demonstrated that the classification model performs very well for the Major Tsunami, Minor Tsunami, and Tsunami classes, achieving high precision, recall, and F1-Score with very high accuracy of 98%. However, the macro average indicates uneven performance across classes, as there are instances of ‘No Warning’ in some coastal gauges. Conclusion: To improve the model performance in the ‘No Warning’ class, it is necessary to balance the dataset by including more ‘No Warning’ scenarios, which can be achieved by simulating additional scenarios involving very small-volume collapse. Additionally, exploring additional collapse parameters such as dip angle and outlier volume could contribute to developing a more robust classification model. Keywords: Machine Learning, Classification, Volcanic Tsunamis, Early Warning, Decision Tree

Robust malicious software detection and classification using global whale optimization algorithm with deep learning approach

Software malware detection and classification leverage sophisticated procedures and methods from the cybersecurity domain for identifying and categorizing malicious software, generally called malware. This procedure analyses code behaviour, file structures, and other features to distinguish between benign and malicious programs. Machine learning (ML) and artificial intelligence (AI) are vital in this domain, allowing the progress of dynamic and adaptive systems that identify novel and developing malware attacks. By training on massive datasets of benign and malicious instances, these systems learn patterns and signatures indicative of malware. This lets them correctly categorize and respond to potential attacks in real-time. This study presents a Global Whale Optimization Algorithm with Neutrosophic Logic for Software Malware Detection and Classification (GWOANL-SMDC) technique. The GWOANL-SMDC technique secures the software via the Android malware recognition process. Primarily, the GWOANL-SMDC technique employs the Neutrosophic Cognitive Maps (NCM) model for the feature selection process. The GWOANL-SMDC technique uses a convolutional long short-term memory (ConvLSTM) model for software malware detection. At last, the GWOA-based parameter tuning is performed to improve the performance of the ConvLSTM model. The simulation values of the GWOANL-SMDC technique are examined on the malware dataset. The obtained results ensured that the GWOANL-SMDC technique improved capability in detecting software malware.

Robust multi-label classification via data reconstruction by neighborhood samples augmentation

Robust multi-label classification via data reconstruction by neighborhood samples augmentation

Differential evolution optimization based ensemble framework for accurate cervical cancer diagnosis

Cervical cancer remains a significant global health concern, demanding accurate and efficient diagnostic approaches for early detection and treatment. However, there are situations where a single deep learning model might not be sufficient to adequately capture the pertinent information required for precise disease prediction from intricate datasets. In this paper, we introduce an innovative ensemble learning framework that leverages the Differential Evolution (DE) algorithm to enhance the detection of cervical cancer. Our innovative approach combines the collective strength of a diverse set of base models using feature fusion techniques to achieve precise and robust classification. We have developed three base models, denoted as Fusion 1, Fusion 2, and Fusion 3, where each fusion model incorporates two pre-trained models using a feature fusion technique. To capture sequential dependencies and improve feature attention, we have integrated ConvLSTM layers and Squeeze-and-Excitation (SE) blocks into these base models. Subsequently, we've amalgamated the predictions of each base model using a weighted aggregation scheme, with the weights optimized using the DE algorithm. Our ensemble framework introduces a novel method for dynamic weight adjustment, utilizing advanced fusion techniques to improve cervical cancer detection accuracy. This approach ensures robust diagnostic performance across cancer cells datasets, enhancing patient outcomes and community health. Our evaluation covers two datasets, SIPaKMeD and Mendeley LBC, demonstrating the remarkable performance and robustness of our ensemble model. It consistently outperforms individual base models and even surpasses state-of-the-art methods, achieving an impressive accuracy rate of 98.96 % on the Mendeley LBC dataset and 95.68 % on the SIPaKMeD dataset. The incorporation of ConvLSTM layers, SE blocks, and DE optimization significantly contributes to the model's exceptional accuracy, its ability to generalize effectively, and its adaptability to diverse scenarios. Additionally, we employ Grad-CAM to visualize how the base models focus on crucial regions for accurate predictions, providing insights into the model's decision-making process. Our ensemble model exhibits outstanding Area under the Curve (AUC) values in Receiver Operating Characteristic (ROC) analysis, further confirming its effectiveness. This research represents a significant advancement in cervical cancer detection, underscoring the potential of ensemble learning techniques optimized through DE for enhancing diagnostic accuracy and advancing patient care.

A Robust and Efficient Hybrid Classification Model for Early Diagnosis of Chest X-Ray Images of COVID-19

لقد كانت هناك جائحة كوفيد-19 منذ ديسمبر 2019، ويتطلب العلاج الطبي الناجح لمرضى كوفيد-19 تشخيصًا سريعًا ودقيقًا. تتطلب مكافحة جائحة كوفيد-19 نظامًا آليًا يستخدم التعلم بالانتقال العميق لتشخيص الفيروس باستخدام الأشعة السينية على الصدر (CXR). يتم استخدام CXR بشكل متكرر في الرعاية الصحية لأنها توفر إمكانية التشخيص السريع والدقيق للأمراض. تتضمن أنظمة التشخيص الآلي بمساعدة الكمبيوتر (CAD) التعلم الآلي أو التعلم العميق لتعزيز الكفاءة والدقة، وبالتالي تقليل المشكلات المستقبلية. يمكن استخدام العديد من أنظمة الذكاء الاصطناعي القائمة على التعلم العميق للتشخيص؛ ومن بين أكثر هذه الشبكات استخدامًا شبكة CNN، التي تم تطويرها لأول مرة وأظهرت دقة مشجعة في تحديد المرضى المؤكد إصابتهم بفيروس كوفيد-19 باستخدام صور CXR. ومن خلال استخدام صور الأشعة السينية، سيعمل هذا العمل على تصميم التعلم الآلي والتعلم العميق لتوفير تشخيص أسرع لعدوى كوفيد-19. ونتيجة لذلك، تستخدم تقنية التعلم بالنقل العميق نموذجًا موجودًا أولاً، ثم تطبق البيانات المطلوبة عليه مرة أخرى. حيث تم استخدام نموذج التعلم Densenet201transfer، وهو أحد تقنيات التعلم عن بعد، كاستخراج الميزة ودمجها مع خوارزمية الإدراك الحسي متعدد الطبقات؛ تم تطبيق هذه التقنية على مجموعة بيانات تابعة للمعهد الوطني للصحة (NIH)، حيث تم استخدام العديد من مقاييس الأداء مثل الدقة والدقة والنوعية والحساسية، حيث أثبتت التجربة كفاءة الخوارزمية المستخدمة من حيث الدقة بنسبة 98.82 %. تعتبر هذه النتائج مشجعة عند مقارنتها بنماذج التعلم عن بعد الأخرى التي تم تدريبها على مجموعة البيانات المماثلة

Current Situation of Non-cleft Craniofacial Anomalies in Thailand: A Multicenter Study.

Non-cleft craniofacial anomalies significantly impact patients' lives and health care systems. This study addresses the gap in the literature concerning these anomalies. Conducted at 4 major hospitals across Thailand, the research aims to provide an overview and understand the experiences of caregivers. The authors conducted a cross-sectional descriptive study in a 3-year period involving 224 primary caregivers of young patients with non-cleft craniofacial anomalies (demographics, medical history, psychosocial aspects). The authors found the highest prevalence of anomalies in the Northern (26.8%) region and the lowest in the Western (3.1%) region. Craniosynostosis and clefts were the predominant anomalies (67.0%), with a surprising lack of family history (7%). Diagnoses were frequently established at regional hospitals (42.0%). Notably, the average outpatient visit was 18.22. A substantial portion of patients required multiple hospital stays: 65.6% with 1 to 5 stays and 9.8% with more than 5 stays. Despite frequent visits, 29% had not received necessary surgery. Of caregivers, 78% had a family income in the middle range (5000-50,000 Thai baht/mo), yet a significant proportion (59.4%) perceived their income as adequate. Their education was bachelor's and higher in only 27.2%. Although stress was common (62.1%), most caregivers (79.0%) did not consider it a burden. The study highlights the challenges faced by both patients and caregivers. It reveals a need for improved access to specialized care, more specialists, dedicated centers, and support networks. A more robust classification system is also essential. To optimize care, a comprehensive and region-specific health care approach is crucial.

Robust Classification via Finite Mixtures of Matrix Variate Skew-t Distributions

Robust Classification via Finite Mixtures of Matrix Variate Skew-t Distributions

Discriminant analysis using feature extraction from spectral domain responses to achieve accurate delineation for robust evaluation or classification of soil properties

ABSTRACT Soil spectroscopy offers a method for quantitatively analysing soil chromophores and employing a screening approach. However, utilizing general models with this data at once doesn’t consistently yield accurate results in determining soil parameters. Ambiguities in response within certain domains of the reflectance spectrum are observed. To address this issue, a method is proposed where quantities associated with spectral response are outputted within well-defined domains corresponding to specific parameters. This article introduces a novel approach centred on identifying features from reflectance spectra to assess soil properties and mitigate issues arising from a heterogeneous soil dataset. The focus is on swiftly automating the classification of soil spectral data by extracting features directly related to the varied composition of chromophores in the samples. These features are categorized for each chromophore parameter and analysed using linear discriminant analysis (LDA). The study employs the ‘309 soil samples’ from the GEO-CRADLE open spectral soil library in Greece. Specifically, relevant data from six spectral responses are extracted to separate and quantify certain soil parameters such as clay content, NO3− levels, and SOM (Soil Organic Matter), despite their interrelations. We developed a method to visually represent the determination outcomes quantitatively by establishing three threshold parameter values: 1.6% for SOM, 19% for clay, and 13.6 ppm for NO3−. These determinations will span the explored ranges of ’0.3 to 4.18%’ for SOM, “3 to 48%” for clay, and ’0 to 661.2 ppm’ for NO3−. Utilizing supervised LDA proves notable classification rates of 93.3% for the SOM task, 90% for the Clay task, and 86.6% for NO3− with the specific cut-off value parameter. The final result is presented as a map showing the positions of various prediction samples, highlighting the percentages of clay and organic matter relative to 19% and 1.6%, respectively, as well as the nitrate concentration in these samples relative to 13.6 ppm.

Cross-Modal Harmony: Low-Rank Fusion for Enhanced Artificial Emotion Recognition

Emotion recognition has become a subject of considerable interest in recent times, owing to its diverse and far-reaching applications in various fields. These applications span from enhancing human-computer interactions to assessing mental health and improving entertainment systems. The proposed study presents a novel approach for emotion recognition by fusing audio and video modalities using low-rank fusion techniques. The proposed methodology leverages the complementary nature of audio and video data in capturing emotional cues. Audio data often encapsulates tone, speech patterns, and vocal nuances, while video data captures facial expressions, body language, and gestures. However, the challenge lies in effectively integrating these two modalities to enhance recognition accuracy. To address the challenge, it employs low-rank fusion, a dimensionality reduction technique that extracts the most informative features from both modalities while minimizing redundancy. Furthermore, it presents the implementation of the chosen low-rank fusion algorithm in a real-world emotion recognition system. The results can contribute to advancing the field of emotion recognition by providing a practical and efficient solution for combining audio and video data to achieve more robust and accurate emotion classification. Keywords: Deep Learning; Emotion Recognition; Human-Computer Interaction; Low-Rank Fusion; Multimodal Fusion.

Machine Learning Ensemble Methodologies for the Prediction of the Failure Mode of Reinforced Concrete Beam–Column Joints

One of the most critical aspects in the seismic behavior or reinforced concrete (RC) structures pertains to beam–column joints. Modern seismic design codes dictate that, if failure is to occur, then this should be the ductile yielding of the beam and not brittle shear failure of the joint, which can lead to sudden collapse and loss of human lives. To this end, it is imperative to be able to predict the failure mode of RC joints for a large number of structures in a building stock. In this research effort, various ensemble machine learning algorithms were employed to develop novel, robust classification models. A dataset comprising 486 measurements from real experiments was utilized. The performance of the employed classifiers was assessed using Precision, Recall, F1-Score, and overall Accuracy indices. N-fold cross-validation was employed to enhance generalization. Moreover, the obtained models were compared to the available engineering ones currently adopted by many international organizations and researchers. The novel ensemble models introduced in this research were proven to perform much better by improving the obtained accuracy by 12–18%. The obtained metrics also presented small variability among the examined failure modes, indicating unbiased models. Overall, the results indicate that the proposed methodologies can be confidently employed for the prediction of the failure mode of RC joints.

Classification of Rice Grains by Image Processing

General Background: The assessment of the surface quality of pre-treated rice grains is crucial for determining their market acceptability, storage stability, processing quality, and overall consumer satisfaction. Traditional evaluation methods are often time-consuming and yield subjective classifications. Specific Background: The lack of a comprehensive tagged image dataset hinders the application of advanced convolutional neural networks (CNN) for detailed damage classification of healthy rice grains. Knowledge Gap: Existing datasets and methods limit the effective exploration of sophisticated CNN models for categorizing rice types, particularly in identifying subtle damage characteristics. Aims: The study aims to create a robust rice grain classification system using image processing techniques, primarily deep learning algorithms, to improve the classification of rice varieties. Results: Utilizing a dataset of 75,000 images across five widely cultivated rice varieties in Turkey, we achieved classification accuracies of 100% for CNN, 99.95% for deep neural networks (DNN), and 99.87% for artificial neural networks (ANN). Novelty: The proposed approach represents a significant advancement in rice classification technology, employing a combination of image acquisition, feature extraction, and machine learning to streamline the process, effectively addressing the challenges faced in traditional methods. Implications: The findings underscore the potential for improved sorting and grading efficiency in the rice industry, facilitating better market outcomes and consumer satisfaction through enhanced quality control Highlights: Innovative: Uses CNN for accurate rice variety classification. Data-Driven: Analyzes 75,000 images for enhanced quality evaluation. Impactful: Increases efficiency in sorting and grading processes. Keywords: Rice Classification, Image Processing, CNN, Machine Learning, Agricultural Technology

Multimodal Fusion of EEG and Audio Spectrogram for Major Depressive Disorder Recognition Using Modified DenseNet121.

This study investigates the classification of Major Depressive Disorder (MDD) using electroencephalography (EEG) Short-Time Fourier-Transform (STFT) spectrograms and audio Mel-spectrogram data of 52 subjects. The objective is to develop a multimodal classification model that integrates audio and EEG data to accurately identify depressive tendencies. We utilized the Multimodal open dataset for Mental Disorder Analysis (MODMA) and trained a pre-trained Densenet121 model using transfer learning. Features from both the EEG and audio modalities were extracted and concatenated before being passed through the final classification layer. Additionally, an ablation study was conducted on both datasets separately. The proposed multimodal classification model demonstrated superior performance compared to existing methods, achieving an Accuracy of 97.53%, Precision of 98.20%, F1 Score of 97.76%, and Recall of 97.32%. A confusion matrix was also used to evaluate the model's effectiveness. The paper presents a robust multimodal classification approach that outperforms state-of-the-art methods with potential application in clinical diagnostics for depression assessment.

A3GT: An Adaptive Asynchronous Generalized Adversarial Training Method

Adversarial attack methods can significantly improve the classification accuracy of deep learning models, but research has found that although most deep learning models with defense methods still show good classification accuracy in the face of various adversarial attack attacks, the improved robust models have a significantly lower classification accuracy when facing clean samples compared to themselves without using defense methods. This means that while improving the model’s adversarial robustness, it is necessary to find a defense method to balance the accuracy of clean samples (clean accuracy) and the accuracy of adversarial samples (robust accuracy). Therefore, in this work, we propose an Adaptive Asynchronous Generalized Adversarial Training (A3GT) method, which is an improvement over the existing Generalist method. It employs an adaptive update strategy without the need for extensive experiments to determine the optimal starting iteration for global updates. The experimental results show that compared with other advanced methods, A3GT can achieve a balance between clean sample classification accuracy and robust classification accuracy while improving the model’s adversarial robustness.

Multi-modal classification of breast cancer lesions in Digital Mammography and contrast enhanced spectral mammography images

Breast cancer ranks as the second most prevalent cancer in women, recognized as one of the most dangerous types of cancer, and is on the rise globally. Regular screenings are essential for early-stage treatment. Digital mammography (DM) is the most recognized and widely used technique for breast cancer screening. Contrast-Enhanced Spectral Mammography (CESM or CM) is used in conjunction with DM to detect and identify hidden abnormalities, particularly in dense breast tissue where DM alone might not be as effective. In this work, we explore the effectiveness of each modality (CM, DM, or both) in detecting breast cancer lesions using deep learning methods. We introduce an architecture for detecting and classifying breast cancer lesions in DM and CM images in Craniocaudal (CC) and Mediolateral Oblique (MLO) views. The proposed architecture (JointNet) consists of a convolution module for extracting local features, a transformer module for extracting long-range features, and a feature fusion layer to fuse the local features, global features, and global features weighted based on the local ones. This significantly enhances the accuracy of classifying DM and CM images into normal or abnormal categories and lesion classification into benign or malignant. Using our architecture as a backbone, three lesion classification pipelines are introduced that utilize attention mechanisms focused on lesion shape, texture, and overall breast texture, examining the critical features for effective lesion classification. The results demonstrate that our proposed methods outperform their components in classifying images as normal or abnormal and mitigate the limitations of independently using the transformer module or the convolution module. An ensemble model is also introduced to explore the effect of each modality and each view to increase our baseline architecture's accuracy. The results demonstrate superior performance compared with other similar works. The best performance on DM images was achieved with the semi-automatic AOL Lesion Classification Pipeline, yielding an accuracy of 98.85 %, AUROC of 0.9965, F1-score of 98.85 %, precision of 98.85 %, and specificity of 98.85 %. For CM images, the highest results were obtained using the automatic AOL Lesion Classification Pipeline, with an accuracy of 97.47 %, AUROC of 0.9771, F1-score of 97.34 %, precision of 94.45 %, and specificity of 97.23 %. The semi-automatic ensemble AOL Classification Pipeline provided the best overall performance when using both DM and CM images, with an accuracy of 94.74 %, F1-score of 97.67 %, specificity of 93.75 %, and sensitivity of 95.45 %. Furthermore, we explore the comparative effectiveness of CM and DM images in deep learning models, indicating that while CM images offer clearer insights to the human eye, our model trained on DM images yields better results using Attention on Lesion (AOL) techniques. The research also suggests a multimodal approach using both DM and CM images and ensemble learning could provide more robust classification outcomes.