Practical Radiation Oncology

Li X.

Image classification is a fundamental task in computer science, underpinning various applications such as object detection, face recognition, and object interaction analysis. The concept holds significant value due to its wide-ranging applications across multiple fields. Traditional methods for image classification, however, have been limited by their slow processing speed, rigidity, and high costs. The integration of deep learning models, particularly Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs), has revolutionized this process, enabling the development of automated, fast, and practical systems. These advanced models are now employed in diverse areas, including biomedical science, remote sensing, and business management, thanks to their ability to achieve high accuracy across a broad spectrum of scenarios. Training these models involves the use of well-known datasets like Canadian Institute for Advanced Research (CIFAR) and Modified National Institute of Standards and Technology (MNIST), which provide the necessary data for optimization and validation. The paper examines the structure, functionality, advantages, and limitations of CNNs and SVMs in the context of image classification, demonstrating that deep learning-driven classification is now a mainstream research focus. This study highlights the transformative impact of these models and provides insights into their future potential.

Chen A.

This study focuses on analysing and exploring the You Only Look Once (YOLO) algorithm. Specifically, this article analyses the evolution and performance of three versions (YOLOv1, YOLOv5, and YOLOv8) in object detection. The research begins by detailing the fundamental concepts of object detection and the datasets commonly used in this field. It then delves into the specific architectures and experimental outcomes associated with each YOLO version. The analysis reveals that while YOLOv8 introduces advanced features and improvements, earlier versions like YOLOv5 may offer superior stability and performance under certain conditions, particularly in specific tasks such as car detection. The discussion emphasizes the significant impact of factors such as batch size on model performance, suggesting that fine-tuning these parameters can optimize the algorithm for particular applications. The study concludes that the future of YOLO development lies in exploring and refining different variants, particularly those of YOLOv8, to better meet diverse requirements. By focusing on five distinct YOLOv8 variants, the research aims to enhance the adaptability and effectiveness of the YOLO framework across a wide range of object detection challenges, thereby contributing valuable insights into the ongoing advancement of this technology.

Jin Z.

Glaucoma, the leading cause of irreversible blindness, must be diagnosed early and thus treated in time. However, it has no noticeable symptoms in its early stages and may not be detected easily. This paper aims to compare two well-known convolutional neural network (CNN) structures, namely Fully Convolutional Networks (FCNs) and U-Net for the segmentation of the optic disc (OD) and optic cup (OC) from retinal fundus images which play an important role in glaucoma diagnosis. The performance of both models is assessed using qualitative parameters such as the Dice coefficient, Jaccard index, and cup-to-disc ratio (CDR) error. In our experiment, the U-Net model yields more accurate segmentation results with 0.9601 average pixel accuracy and 0.9255 dice score for OD segmentation, outperforming the FCNs model with 0.9560 average pixel accuracy and 0.9132 dice score for OD segmentation. However, FCNs have a shorter inference time of 0. 0043 seconds against U-net’s 0. 0062 seconds making FCNs more suitable for real-time applications. The restrictions related to this study include biases from using only one dataset acquired from particular imaging devices, dependency on mask-based cropping techniques, and comparison being restricted to two fundamental architectures. This work presents the contribution of the deep learning models in improving glaucoma screening and therefore helping in avoiding blindness.

Wang Z.

Consumer reviews are an important source of data used to judge and examine consumer sentiment, and data mining for reviews of electronic products is an important way to help improve the design of electronic products. The research is based on the consumer reviews of online cell phone e-commerce, The paper constructs a sentiment dictionary in this field based on the Sentiment Oriented Point Mutual Information (SO-PMI) algorithm, and the sentiment weight of the review word vectors. An extreme Gradient Boosting Tree (XGBoost) is used to integrate word vectors and a Large Language Model (LLM) to construct a sentiment recognition model, and finally, a review sentiment index is derived, which unfolds from multiple dimensions to analyze the sentiment tendency in consumer reviews. The empirical analysis shows that the accuracy, recall, area under the curve (AUC), and other validation indexes of the constructed sentiment recognition model are further improved compared with the LLM model, which has a certain application value. When applying the weighted word vector method, the model has been significantly improved compared with the LLM model, the accuracy is increased by 5%, the accuracy is increased by 10%, and the comprehensive accuracy is increased by 2% after the comprehensive application of the two.

Chen Z.A.

The 2022 FIFA World Cup final attracted 1.5 billion viewers, while billions of dollars are wagered on soccer matches every year. The increasing demand for accurate predictions, both for academic research and betting purposes, has driven the development of advanced forecasting models. This study explores the application of mathematical and machine learning models to predict results of soccer matches, with the dual aim of academic advancement and profitable betting. The author utilizes a comprehensive dataset from top European leagues (2014-2022) and employ models including Bivariate Poisson Distribution, Naive Bayes, Neural Networks, Support Vector Machines, Random Forests, and Gradient Boosting. The paper’s feature engineering combines historical match statistics, FIFA ratings, and betting odds. While Random Forests achieved the highest accuracy (56.25%), predicting draws remains challenging. The study highlights the potential for improved prediction systems and suggests future research in advanced draw prediction techniques and profitability analysis, the paper provides research directions for researchers in related fields.

Liu Y.

Predicting traffic flow is crucial for the functionality of intelligent transportation systems. It is of critical importance to relieve traffic pressure, reduce accident rates, and alleviate environmental pollution. It is an important part of the construction of modern intelligent road networks. With advancements in deep learning (DL), DL models have made notable strides in prediction. However, due to the complexity and non-transparency of DL models themselves, there are still problems of low accuracy and interpretability in traffic flow prediction (TFP). Leveraging large language models (LLM) helps to improve the negative conditions caused by other DL models in prediction. This paper first briefly summarizes the basic characteristics of LLM and their advantages in TFP; then conducts relevant research and analysis in the order of experimental design steps comparison and results and conclusions comparison; then analyzes and discusses the current problems and challenges faced by LLM; finally, it looks forward to future research directions and development trends, and summarizes this paper.

Li S., Zhang B.

Hierarchical Federated Learning (FL) presents a novel approach that combines global model training with localized personalization fine-tuning to enhance the predictive accuracy of decentralized machine learning systems. Traditional FL methods, which allow multiple clients to collaboratively train a global model without sharing raw data, are hindered by issues such as non-independent and identically distributed (non-IID) data, communication overhead, and limited generalization across diverse client datasets. This study proposes a hierarchical model that mitigates these challenges by incorporating a global model, trained using the Federated Averaging (FedAvg) algorithm, and applying client-specific fine-tuning to improve local model performance. The experiment conducted on a movie recommendation system demonstrates that this hierarchical approach significantly reduces the global model’s error while offering personalized improvements on client-specific datasets. Results show an average Root Mean Squared Error (RMSE) reduction of 0.0460 following local personalization. This hybrid approach not only enhances model accuracy but also preserves data privacy and increases scalability, making it a promising solution for decentralized recommendation systems.

Bao Y.

As heart disease has become the leading cause of death worldwide, early and accurate prediction is crucial to help doctors make initial judgments about patients and improve their survival rates. This study aims to improve the accuracy and efficiency of heart disease prediction through Machine learning (ML) methods to help medical diagnosis. A heart disease dataset was used in the study, and multiple ML models were used to analyze multiple key health features, and the model performance was verified through a test set. This paper concludes that Logistic regression and random forests perform well in this task and have high practical value. Future research can stack models and optimize data sources to improve the practical performance of the model. This study provides a basic framework for building an intelligent medical auxiliary diagnosis system, which helps to achieve early prevention and timely judgment of heart disease, thereby improving the overall efficiency of medical services.

Shi L.

This paper is an in-depth study by delving into the latest in image generation technology, where thesis is focusing on the Generative Adversarial Networks (GANs) and MirrorGANs possibilities. Image Generation is the backbone of visual computing, mostly utilized in intelligent designs. It is for this reason that this research aims at unravelling the theoretical basis and consolidated practices of GANs when it conies to generating both high-quality and semantically consistent imagery. The study will investigate the whole of the image generation process, starting from data preprocessing to the use of GANs to generate images from textual descriptions. The work discussed the relevance as well as the limitations of these technologies from the artistic point of view, medical imaging, and virtual reality. Tire article concludes that the paper sketches the data and experiments that show that the realism and richness hi picture quality are accentuated when GANs and MirrorGANs are incorporated. This suggests the scope of image-generation technology to enhance human-machine collaboration and allow for innovating hi smart tech. Further studies will be geared to enhancing these methods and consequently drawing humanity and machines closer, which hi nun will fuel the ongoing progress in this fast-paced sphere.

Zhang M.

Security in the World Wide Web has recently seen an enormous upgrade in almost every aspect. Identifying malicious activities hi a network such as network attacks and malicious users plays a significant role hi these upgraded security directions. This research utilizes the KDDCUP99 dataset to incorporate K-means clustering with three classifiers: Random Forest (RF). Naïve Bayes (NB). and Support Vector Machine (SVM) with the goal to boost the accuracy of predicting network intrusions. In tins paper. K-means clustering technique is applied as a preprocessing step to enhance the overall quality of network intrusion detection and maximize the accuracy of the network security measures. The goal is to identify anomalies with high accuracy. Experimental results hidicate that the optimal combination is K-means + RF. which outperformed the others hi precision, recall, and Fl-score. Although K-means + NB demonstrated superior recall for certahi smaller anomalies, it underperformed compared to the RF model. The paper concludes by highlighting the value of ensemble approaches, in particular Random Forest, for tackling anomaly detection and network security issues, particularly hi light of the expanding significance of social networks and the internet.

Yang R.

Air pollution forecasting for public health and policy-making has a critical importance, this paper employs a Long Short-Term Memory (LSTM) model to perform in-depth prediction of PM2.5 concentrations measured at the U.S. Embassy in Beijing, outperforming regular forecasting approaches. In the LSTM model, the research examines a very detailed hourly dataset and beats regular forecasting approaches. A key finding is the model’s ability to effectively generalize from historical data to predict future air quality trends, with its adeptness at handling time-dependent relationships. This research emphasizes the importance of LSTM in air pollution prediction and management in environmental science as it provides an effective means for planning and making decisions on air quality management. This research is of great importance in providing a groundwork for further enhancement of prediction modeling. By offering a more reliable and sophisticated picture of air quality variations, this study addresses the current problem about how urban air pollution control could be improved in the city.

Ren C.

In the era of rapid advancements in natural language processing (NLP) models, these technologies have immense potential to detect and address societal issues, enhancing the functioning of the digital society. Online shopping platforms rely heavily on user reviews to influence buyer decisions, yet malicious reviews can significantly degrade user experience. This study focuses on analyzing the emotional polarity of malicious brushorder (falsely generated) reviews in e-commerce product comments, utilizing the Jingdong product review dataset. The methodology involves utilizing the Word2Vec model to vectorize the text data, followed by principal component analysis (PCA) for outlier detection to identify potential malicious reviews based on their unique characteristics. The PCA results are further leveraged for dimensionality reduction, simplifying the dataset. Subsequently, the BERT model is employed to perform semantic similarity analysis, allowing for the screening and expansion of the experimental dataset with similar malicious comments. This enriched dataset is then subjected to sentiment polarity analysis, enabling a deeper tinderstanding of the nature and impact of these malicious reviews. By facilitating buyers in making informed decisions based on genuine reviews, this research underscores the practical value of NLP hi addressing real-world challenges in e-commerce.

Ye J.

In recent years, research on Natural Language Processing (NLP) has made consistent progress and has become a popular topic. As a promising branch of Machine Learning, NLP focuses on the understanding, generating and analysing of human languages. The applications of NLP include chatbots and language translation. This paper represents a HanLP based NLP model. The model is capable of analysing the literary style of given Chinese text by quantifying the literary style of the text on the basis of five fundamental elements, namely literary grace, sentiments, momentum, climate and lingering charm. This paper presents the input and output data of the research and conducts analyses on these data. Moreover, this paper draws a conclusion on the deviation rate and robustness of the model. It is reckoned that this model initially possesses the function of literary style analysis of Chinese text. The research, per se, along with its data, is capable of being reference for research in NLP and related fields.

Lu Z.

A pervasive security issue in web applications is database injection, enabling attackers to alter SQL queries in order to get unauthorized access to confidential information. Using the Naive Bayes classifier, a probabilistic model specifically developed for text classification tasks, this work introduces a novel method for detecting SQL injection vulnerabilities.The process begins by collecting and organizing a comprehensive dataset, which includes both harmful and non-malicious SQL queries. Feature extraction is later employed to identify patterns and characteristics commonly associated with SQL injection, such as certain SQL clauses and logical operators. This collection of attributes is employed to generate a feature vector that serves as the input for the Naive Bayes classification algorithms. The classifier is trained using a labeled dataset and then learns to distinguish between benign and malicious requests by assessing their computed probabilities. Conventional measures such as accuracy, precision, recall, and F1-score are employed to assess the model’s ability in correctly identifying SQL while reducing false positive classifications.The present study demonstrates the potential of Naive Bayes in enhancing online application security by providing a methodical and scalable strategy for identifying SQL injection attacks.

Zhang H.

In modern society, the restoration of artwork has become increasingly important. Generative models can provide reference images for the damaged or blurred core areas of these artworks. This paper simulates artificial damage to classic portrait paintings in the Art Portraits dataset by adding center masks during data preprocessing and then implements the image inpainting task. During the training phase, the Denoising Diffusion Probabilistic Model (DDPM) is fine-tuned by progressively adding noise to the center-masked images in the noising stage, followed by denoising in the denoising stage to generate images. The generated images are compared with the original undamaged images through loss calculations to optimize the model. Additionally, a Generative Adversarial Network (GAN), which has shown promising results on other datasets, is used as a baseline for comparison. The damaged images are used as inputs, and the generated images are compared to the ground truth to evaluate the performance of both models. In the testing phase, two widely used metrics in image evaluation, Mean Squared Error (MSE) and Fréchet Inception Distance (FID), are introduced to assess the performance. The fine-tuned DDPM achieves an MSE of 0.2622 and an FID of 16.85, while the GAN scores 0.2835 and 22.78, respectively. Since lower values indicate higher fidelity in reproducing the original image, which is crucial for art restoration, the conclusion drawn from this paper is that the fine-tuned DDPM demonstrates higher accuracy and is more suitable for restoration projects related to Art Portraits.