Malicious websites present a substantial threat to the security and privacy of individuals using the internet. Traditional approaches for identifying these malicious sites have struggled to keep pace with evolving attack strategies. In recent years, language models have emerged as a potential solution for effectively detecting and categorizing malicious websites. This study introduces a novel Bidirectional Encoder Representations from Transformers (BERT) model, based on the Transformer encoder architecture, designed to capture pertinent characteristics of malicious web addresses. Additionally, large-scale language models are employed for training, dataset assessment, and interpretability analysis. The evaluation results demonstrate the effectiveness of the large language model in accurately classifying malicious websites, achieving an impressive precision rate of 94.42%. This performance surpasses that of existing language models. Furthermore, the interpretability analysis sheds light on the decision-making process of the model, enhancing our understanding of its classification outcomes. In conclusion, the proposed BERT model, built on the Transformer encoder architecture, exhibits robust performance and interpretability in the identification of malicious websites. It holds promise as a solution to bolster the security of network users and mitigate the risks associated with malicious online activities.

In histopathological image segmentation, existing methods often show low sensitivity to small lesions and indistinct boundaries, leading to missed detections. Since, in clinical diagnosis, the consequences of missed detection are more serious than false alarms, this study proposes ClinSegNet, a recall-oriented and human-centred framework for reliable histopathology screening. ClinSegNet employs a composite optimisation strategy, termed HistoLoss, which balances stability and boundary refinement while prioritising recall. An uncertainty-driven refinement mechanism is further introduced to target high-uncertainty cases with limited fine-tuning cost. In addition, a clinical data processing pipeline was developed, where pixel-level annotations were automatically derived from IHC-to-H&E mapping and combined with public datasets, enabling effective training under limited clinical data conditions. Experiments on the NuInsSeg and NuInsSeg-UHS datasets showed that ClinSegNet achieved recall scores of 0.8803 and 0.8917, further improved to 0.8983 and 0.9053 with HITL refinement, while maintaining competitive Dice and IoU. Comparative and ablation studies confirmed the complementary design of the framework and its advantage in capturing small or complex lesions. In conclusion, ClinSegNet provides a clinically oriented, recall-prioritised framework that enhances lesion coverage, reduces the risk of missed diagnosis, and offers both a methodological basis for future human-in-the-loop systems and a feasible pipeline for leveraging limited clinical data.