The development of lung cancer poses a severe threat to the lives of the population, making early detection necessary. The use of CT imaging to investigate pulmonary nodules and to conduct CAD is a widespread practice in the field of medicine. In this study, an imbalanced area-under-the-curve (AUC) particle swarm support vector machine (ImAUC-PSVM) optimized by a differential evolution (DE) algorithm has been proposed. The method tackles two significant problems of the previous works related to lung cancer detection: class imbalance and sensitivity to hyperparameters. ImAUC-PSVM incorporates AUC maximization in the cost function, makes training easier, and is appropriate for imbalanced medical data. In addition, the model retains the basic structure of the PSVM, which allows for efficient retraining in clinical applications. To increase the optimization efficiency, the K-Means Clustering (KMC)-based mutation operator was introduced to the DE algorithm. In this approach, the KMC helps to find promising groups of the candidate solutions, and the updating operator proposes new candidates to the current population. The proposed model showed excellent results on the LIDC-IDRI and LUNA-16 databases, reaching F-measure of 91.256
The exponential growth of online textual content, including extensive collections of news articles, scientific research papers, legal documents, and other domain-specific materials, has increased the need for effective automatic text summarization. It aims to generate concise and accurate summaries, thereby reducing the time and effort required to process and comprehend large volumes of textual information. An efficient CAViaR Wave Search Algorithm-optimized Hierarchical Parallel Convolutional Forward Harmonic Network (CWSA-HPCFH-Net) is proposed in this work to enhance the performance of abstractive summarization. The text documents obtained from the dataset are used as input and tokenized using the Bidirectional Encoder Representations from Transformers (BERT) tokenizer, which converts the text into a sequence of tokens. Next, extractive summarization is performed employing two techniques, namely the proposed CWSA-based HPCFH-Net and the graph-based framework. Initially, features are processed through the HPCFH-Net to generate sentence scores. Here, the HPCFH-Net is tuned utilizing the proposed CWSA, which integrates the Conditional Autoregressive Value at Risk (CAViaR) and Wave Search Algorithm (WSA). Secondly, a graph-based framework is utilized to generate sentence scores. Next, fuzzy logic is employed to fuse the two sets of sentence-ranking scores and establish a unified importance score for each sentence. The selected sentences form the extractive summary, which subsequently serves as the input to GPT-NeoX for producing the final abstractive summary. The devised CWSA-based HPCFH-Net achieved superior precision, recall, F-measure, and rouge of 92.674
Text summarization and text classification are key tasks in natural language processing (NLP) that aim to extract meaningful insights from large volumes of text. The goal of text summarization is to condense lengthy content while preserving essential information, whereas text classification involves assigning text to specific topics or labels. This paper proposes a novel method for text summarization and classification, designed to improve efficiency and accuracy by combining advanced deep learning techniques. The process begins with preprocessing the input text using the BERT tokenization model, which converts raw text into tokenized sequences while preserving semantic relationships between words. Following this, various features are extracted from the preprocessed text, including sentence features, Senti-word net features and an improved version of Term Frequency-Inverse Document Frequency (Im.TF-IDF). These features are then assigned importance scores, which are ranked to determine the most relevant portions of the text to include in the summary. Based on the ranked features, an attention-based Long Short-Term Memory (Attn-LSTM) model is utilized to create a coherent and short summary by focusing on the most significant sentences. After summarization, the summarized text is classified by extracting semantic and similarity features, which are fed into an Improved Parallel Convolutional Neural Network (IPCNN). The IPCNN then classifies the text into predefined categories with high accuracy. This approach ensures the retention of essential information in the summary while improving the complete performance of text classification tasks. The IPCNN model attained the highest precision of 0.933, F-measure of 0.929 and Accuracy of 0.971, outperforming the traditional approaches.
Radiologists interpret mammograms together with patient information such as age, breast density, and hormonal status. Many automated methods use images alone or combine image and clinical information in the same way for every patient. This may limit their usefulness when clinical records are incomplete. This study examined whether adjusting the contribution of image and clinical information for each patient could improve breast cancer classification. The proposed framework uses separate branches for mammograms and clinical information. It assigns patient-specific weights to the two sources before making the final prediction. Missing clinical values are identified, and the output is calibrated to provide more reliable probabilities. The study included 4415 cases from three cohorts and compared the framework with two single-source models and eight existing fusion methods. Evaluation used patient-level data partitions, a held-out test set, and five repeated runs. The framework achieved areas under the ROC curve of 0.947, 0.887, and 0.958 across the three cohorts, exceeding the image-only model by 0.029, 0.066, and 0.027, respectively. The improvement remained statistically significant in two cohorts after correction for multiple comparisons. Differences from cross-modal attention were 0.004 or less and were not statistically significant. The expected calibration error for the private cohort decreased from 0.070 to 0.026. As clinical information was removed, the framework maintained performance comparable to that of the image-only model, whereas simple concatenation fell below it. These findings support patient-specific combination of mammographic and clinical information, particularly when clinical records are incomplete.
The identification of prevalent information diffusion patterns, which refer to the recurring temporal trajectories through which content propagates across social networks—such as steady growth, initial bursts, or delayed bursts—is crucial for understanding and managing the spread of content, ideas, and influence in social media ecosystems. Current models for this task primarily rely on network topology or historical data, often overlooking the critical influence of socio-psychological factors on user behavior, which limits their explanatory power for complex dissemination dynamics. To bridge this gap, this study proposes a gated ensemble model that integrates quantitative socio-psychological features. We operationalize the primacy effect and conformity psychology through quantifiable behavioural proxies and introduce an adaptive gating mechanism to dynamically fuse representations from graph convolutional networks, graph attention networks, and transformers into a unified feature space for three-class pattern classification. The model demonstrates robust performance, achieving an accuracy of 77.08
