Wufeng Liu

In high-energy physics experiments, the accurate differentiation of subatomic particles is crucial. However, traditional particle identification methods exhibit significant limitations when it comes to handling high-dimensional data, complex features and data labeling. Consequently, there is a pressing need for more advanced techniques to enhance identification accuracy. While many researchers have employed supervised learning algorithms in particle physics experiments, the label generation process is often time-consuming. Therefore, we opted to utilize unsupervised learning methods for direct classification of the collected data. This paper presents a novel unsupervised classification model, DeepSVD-GMM, which integrates singular value decomposition, deep neural networks and Gaussian mixture models. In the experimental results section, due to the lack of existing literature on the application of unsupervised learning in the classification of J/[ ] and Upsilon, we compared our proposed method with classical unsupervised classification models. The results indicate that our method significantly outperforms other models in terms of classification accuracy. This advancement provides new insights and tools for data analysis in high-energy physics experiments, showcasing the significant potential of advanced feature extraction and unsupervised classification techniques.

To address the growing demand for intelligent automation in particle physics analysis, this study proposes K-PINN, a novel hybrid architecture integrating Kolmogorov—Arnold networks (KANs) and physics-informed neural networks (PINNs) for high-precision classification of Z boson decay channels ([ ] and [ ] ee) in high-energy physics experiments. Key innovations include the synergistic combination of B-spline-based interpretable feature extraction via the KANs module with rigorous physics constraints from the PINNs module, specifically enforcing four-momentum conservation and invariant mass-shell conditions. These components are unified through a physics-guided loss function with adaptive weighting to optimize classification fidelity. Evaluated on the CERN open dataset, K-PINN achieves state-of-the-art performance with 96.3% classification accuracy and an F1-score of 0.963, demonstrating a significant improvement over the traditional threshold-based approach. The architecture’s modular design enables direct generalization to other quantum chromodynamics processes through parameter space reconfiguration, establishing an optimal balance between physical interpretability and model performance. This work provides a versatile framework for intelligent particle physics analysis systems, with immediate applications extending to Higgs boson studies and beyond, while offering critical technical infrastructure for next-generation collider experiments.

Research on Z boson decay patterns helps validate Standard Model predictions and explore new physical phenomena. However, analyzing large datasets from proton–proton colliders presents significant challenges, as traditional methods often struggle with efficiency and accuracy. To address this, we propose CNNBoosting, a novel ensemble learning algorithm that combines multiple weak convolutional neural networks (CNNs) to enhance classification performance. CNNBoosting targets two specific Z-boson decay modes (Zee and Zmumu), improving stability and accuracy in classifying particle interactions. Using a Z-boson dataset from Kaggle, we train and evaluate CNNBoosting, achieving impressive results: 92.74% accuracy, 92.74% precision, 92.78% recall and 92.75% F1 score. Comparisons with existing classifiers, including KNN, Random Forest, XGBoost, MLP, SVM,VGG-16, ResNet-50 and Decision Tree, demonstrate the superior performance of CNNBoosting. This is the first application of a CNN model to this dataset, and our results confirm its effectiveness in classifying Z-boson decay patterns, offering an efficient tool for researchers.

In high-energy physics experiments, traditional data analysis methods increasingly show limitations when handling large-scale datasets, highlighting the urgent need for a transition to automated and intelligent analysis techniques to improve accuracy and efficiency. To address this challenge, we utilized an open dataset from CERN and proposed a novel Z-Bagging model, an ensemble learning approach designed to distinguish between two decay processes of the Z boson: Z [ ]+[ ] (Zmumu) and Z [ ] e+e[ ] (Zee). Initially, we tested several classical machine learning models to evaluate their performance in distinguishing these two decay modes. While these models showed some improvements in accuracy, further optimization was identified. Consequently, we introduced the Z-Bagging model, which combines multiple weak learners, such as Random Forest, Support Vector Machine and K-Nearest Neighbors, into a single strong learner. By integrating diverse models, Z-Bagging achieves more accurate and stable predictions. Experimental results demonstrate that the Z-Bagging model significantly outperforms existing models in classification accuracy for distinguishing Zmumu and Zee decay processes. This improvement provides a more precise tool for data analysis in high-energy physics.

Medical image segmentation is an essential process that facilitates the precise extraction and localization of diseased areas from medical pictures. It can provide clear and quantifiable information to support clinicians in making final decisions. However, due to the lack of explicit modeling of global relationships in CNNs, they are unable to fully use the long-range dependencies among several image locations. In this paper, we propose a novel model that can extract local and global semantic features from the images by utilizing CNN and the visual transformer in the encoder. It is important to note that the self-attention mechanism treats a 2D image as a 1D sequence of patches, which can potentially disrupt the image’s inherent 2D spatial structure. Therefore, we utilized the structure of the transformer using visual attention and large kernel attention, and we added a residual convolutional attention module (RCAM) and multi-scale fusion convolution (MFC) into the decoder. They can help the model better capture crucial features and fine details to improve detail and accuracy of segmentation effects. On the synapse multi-organ segmentation (Synapse) and the automated cardiac diagnostic challenge (ACDC) datasets, our model performed better than the previous models, demonstrating that it is more precise and robust in multi-organ medical image segmentation.

Rice is a necessity for billions of people in the world, and rice disease control has been a major focus of research in the agricultural field. In this study, a new attention-enhanced DenseNet neural network model is proposed, which includes a lesion feature extractor by region of interest (ROI) extraction algorithm and a DenseNet classification model for accurate recognition of lesion feature extraction maps. It was found that the ROI extraction algorithm can highlight the lesion area of rice leaves, which makes the neural network classification model pay more attention to the lesion area. Compared with a single rice disease classification model, the classification model combined with the ROI extraction algorithm can improve the recognition accuracy of rice leaf disease identification, and the proposed model can achieve an accuracy of 96% for rice leaf disease identification.

The automatic segmentation of the lung region for chest X-ray (CXR) can help doctors diagnose many lung diseases. However, extreme lung shape changes and fuzzy lung regions caused by serious lung diseases may incorrectly make the automatic lung segmentation model. We improved the U-Net network by using the pre-training Efficientnet-b4 as the encoder and the Residual block and the LeakyReLU activation function in the decoder. The network can extract Lung field features efficiently and avoid the gradient instability caused by the multiplication effect in gradient backpropagation. Compared with the traditional U-Net model, our method improves about 2.5% dice coefficient and 6% Jaccard Index for the two benchmark lung segmentation datasets. Our model improves about 5% dice coefficient and 9% Jaccard Index for the private lung segmentation datasets compared with the traditional U-Net model. Comparative experiments show that our method can improve the accuracy of lung segmentation of CXR images and it has a lower standard deviation and good robustness.

Pneumothorax is a potentially life-threatening disease that requires urgent diagnosis and treatment. Clinically, a chest X-ray examination is the first choice for diagnosing pneumothorax. However, it is difficult to diagnose pneumothorax by only frontal chest X-ray imaging when the lesion area is only composed of a small amount of air. Therefore, we propose a pneumothorax diagnosis neural network based on feature fusion, where frontal and lateral X-ray information are fused. In this network, there are two inputs and three outputs. The two inputs are the frontal chest X-ray image and the lateral chest X-ray image. The three outputs are the classification results of the frontal chest X-ray image, the classification results of the lateral chest X-ray image, and the classification results integrating the characteristics of the fused frontal chest X-ray image and lateral chest X-ray image. Our algorithm considers the vanishing gradient problem in the pneumothorax recognition model and introduces the residual block to alleviate this problem. Because of the large number of channels in this model, we also utilize channel attention mechanisms to improve the model’s performance. Our comparative experiments show that neural network fusion of frontal and lateral chest image features can achieve higher accuracy than the single task model. Using only image-level annotation, our pneumothorax model can achieve high recognition accuracy.

Automatic and highly accurate lung segmentation in chest X-ray (CXR) images is the basis of computer-aided diagnosis systems, because the lung is the region of interest of many diseases, and it can show useful information through its contours. However, automatic lung segmentation is immensely challenging due to extreme variations in the shape, obscure lung area, or opacity caused by lung diseases reaches high-intensity values. In the face of these severe situations, the model may segment the lung boundary incorrectly. We designed an improved U-Net network: using the pre-training Efficientnet-b4 as the encoder, and the residual block and LeakyRelu activation function are used in the decoder. The network can not only extract features with high efficiency but also avoid the gradient explosion caused by the multiplication effect in gradient backpropagation. We constructed a CXR lung field segmentation dataset (Haut) based on the NIH CXR dataset. In particular, this lung segmentation dataset contains some serious abnormal cases, such as lung deformation, pleural effusion, covered by foreign matters, or CXR blur caused by severe lung disease. The improved U-Net is evaluated on Haut, JSRT, and Montgomery County (MC) datasets. Experimental results show that our network can achieve high-precision lung segmentation.