Thyroid nodules are increasingly detected in clinical practice, and accurate imaging evaluation is essential for risk stratification, treatment planning, and postoperative surveillance of thyroid cancer. Although ultrasound remains the first-line modality for thyroid nodule assessment, the roles of computed tomography (CT), spectral CT, magnetic resonance imaging (MRI), and positron emission tomography-computed tomography (PET-CT) vary across clinical scenarios, and the optimal integration of these modalities remains insufficiently standardized. In particular, existing studies often focus on individual imaging techniques, while practical guidance on modality selection for initial screening, preoperative anatomical assessment, recurrence monitoring, and high-risk disease evaluation remains limited. This narrative review summarizes the imaging principles, clinical indications, diagnostic value, advantages, and limitations of ultrasound, CT, spectral CT, MRI, and PET-CT in the evaluation of thyroid nodules and thyroid cancer. Future advances should focus on standardized multimodal imaging strategies, quantitative functional imaging, radiomics, and carefully validated artificial intelligence-assisted approaches to improve individualized diagnosis and management of thyroid nodules and thyroid cancer.

Steatotic donor livers are highly susceptible to post-transplant dysfunction; however, the underlying mechanisms remain incompletely understood. This study aimed to investigate the role of galectin-3 (LGALS3)-mediated pyroptosis in steatotic liver graft injury and explore its therapeutic potential.

A mouse model of steatotic liver transplantation was established. Graft tissues were subjected to RNA sequencing to identify key regulators. In vitro, LGALS3 was modulated in steatotic hepatocytes under ischemia/reperfusion stress to assess its impact on the NLRP3 inflammasome and pyroptosis. The regulatory mechanism by which LGALS3 modulates NLRP3 ubiquitination was further examined. Finally, the therapeutic efficacy of LGALS3 inhibition was evaluated in an orthotopic liver transplantation model.

Transcriptomic analysis identified LGALS3 as a key upregulated molecule in steatotic grafts, associated with pyroptosis pathways. In vitro, LGALS3 overexpression enhanced NLRP3 inflammasome activation and pyroptotic cell death, whereas LGALS3 knockdown exerted protective effects. Mechanistically, LGALS3 modulated NLRP3 inflammasome activity by regulating its ubiquitination. In vivo, pharmacological inhibition of LGALS3 significantly improved graft function, reduced histological injury, suppressed pyroptosis, and prolonged recipient survival.

This study demonstrates that LGALS3 drives steatotic graft injury by promoting NLRP3-mediated pyroptosis through the regulation of ubiquitination. These findings identify LGALS3 as a promising therapeutic target for improving the outcomes of liver transplantation using steatotic donor organs.

The current criterion of biochemical response to ursodeoxycholic acid in primary biliary cholangitis is an alkaline phosphatase (ALP) level of ≤1.67 × the upper limit of normal (ULN) after 12 months of treatment. However, a proportion of patients who meet this parameter may still progress to liver decompensation. This study aimed to optimize the clinical management of primary biliary cholangitis by (1) establishing ALP normalization as a core treatment target, (2) identifying early intervention windows, and (3) developing risk stratification criteria.

This multicenter retrospective study included an internal cohort and an external validation cohort. We assessed the prognostic impact of ALP normalization with Kaplan-Meier and Cox regression. Sankey diagrams and segmented Poisson regression analysis mapped dynamic risk transitions to identify critical intervention windows. Predictive performance (sensitivity/specificity/positive predictive value/negative predictive value [NPV]) of Mayo, Paris II, and Toronto criteria for 12-month ALP normalization was compared.

Patients achieving ALP normalization showed significantly higher complication-free survival versus those with ALP 1.0–1.67 × ULN (89.8% vs. 79.8%; P = 0.016). Segmented Poisson regression identified significant change points at 3.73 and 5.5 months for high-to-medium and medium-to-low risk transitions, respectively. Failure to meet the Toronto criteria at month 3 predicted non-normalization with 95% NPV, whereas Paris II criteria at month 6 provided optimal specificity (73%) for identifying patients who failed to achieve ALP normalization.

ALP normalization significantly improves clinical outcomes. Two subgroups demonstrate low normalization probability and warrant early intervention: (1) patients with ALP ≥ 1.67 × ULN after 3 months and (2) those not meeting Paris II criteria by month 6.

Sepsis is a life-threatening syndrome associated with high morbidity and mortality, underscoring the urgent need for early diagnostic biomarkers and therapeutic targets. However, current diagnostic strategies remain insufficiently precise because of the complex immune dysregulation and immune microenvironment heterogeneity that characterize sepsis. This study aimed to identify reliable diagnostic biomarkers for sepsis and explore their immune regulatory mechanisms together with potential therapeutic relevance using multidimensional bioinformatic analyses.

Single-cell transcriptomic and bulk RNA sequencing datasets were integrated to screen candidate diagnostic genes for sepsis. Immune infiltration, co-expression network and pathway enrichment analyses were performed to explore immune regulatory mechanisms. Machine-learning approaches were used to validate the diagnostic signature, and molecular docking was conducted to predict candidate targeted compounds.

A total of 346 differentially expressed genes were identified and were mainly enriched in immune, coagulation, and metabolic pathways. CIBERSORT and single-cell analyses revealed increased neutrophils, monocytes, and γδ T cells and reduced CD8+ T cells and resting natural killer cells. Four diagnostic genes (S100A12, CD22, CSTA, and UPP1) were prioritized. The four-gene model showed robust external performance (area under the receiver operating characteristic curve = 0.860; sensitivity = 0.781; specificity = 0.780), and interpretability analysis highlighted UPP1 and S100A12 as dominant predictors. Molecular docking suggested potential interactions between these targets and anti-inflammatory compounds.

This integrative framework identifies four immune-related diagnostic genes for sepsis and links them to immune-cell remodeling and candidate therapeutic interactions, providing a basis for future mechanistic and clinical validation.

The optimal surgical management for spontaneous supratentorial intracerebral hemorrhage (SSICH) remains controversial because conventional approaches often fail to balance rapid decompression with effective hematoma evacuation. This study aimed to evaluate the efficacy and safety of new combined surgical strategies (“two-in-one” and “three-in-one”) versus conventional methods for SSICH.

This retrospective cohort study included 451 SSICH patients treated between January 2019 and December 2023. Based on clinical severity, patients were stratified into Group I (non-herniation, n = 374) and Group II (herniation, n = 77). Within each subgroup, patients were further categorized by treatment period: a historical control cohort (2019–2020) receiving conventional surgery, and an intervention cohort (2021–2023) receiving combined strategies (“two-in-one” for Group I; “three-in-one” for Group II). Outcomes included decompression time, hematoma evacuation rate, complications, and six-month functional recovery (Glasgow Outcome Scale/modified Rankin Scale), were compared.

In Group I, the “two-in-one” strategy achieved faster decompression (4.65 min) and a high evacuation rate (92.15%), which was comparable to neuroendoscopy alone (90.58%) and significantly higher than stereotactic aspiration alone (44.55%). This was associated with improved six-month outcomes (poor outcome rates were 39.39%, 54.35%, and 42.86% in Groups I-A, I-B, and I-C, respectively, overall P = 0.034). In Group II, the “three-in-one” strategy demonstrated shorter decompression time (4.73 vs. 37.85 min, P < 0.001) and higher evacuation rates (80.51% vs. 63.50%, P < 0.001) than decompressive craniectomy alone. Logistic regression further supported the prognostic advantage of the “two-in-one” strategy in Group I.

These combined strategies may integrate the advantages of multiple techniques to enable rapid decompression and effective hematoma clearance in SSICH. Prospective studies are warranted.

Colorectal polyp detection from endoscopic images is critical for the early diagnosis of colorectal cancer. However, traditional deep learning methods often suffer from limited generalization when deployed across datasets containing different polyp morphologies. This work aimed to investigate whether vision-language foundation models can facilitate zero-shot generalization across multiple polyp datasets without target-domain fine-tuning.

We introduced a zero-shot colorectal polyp detection framework based on Contrastive Language-Image Pretraining (CLIP) to improve cross-dataset detection performance. Key innovations include: (1) a background patch contrastive loss using pseudo-normal tissue patches to teach the model to distinguish normal mucosa from polyps; (2) attribute-enhanced text prompts that incorporate domain-specific descriptors of polyp appearance, improving the model’s semantic generalization to novel polyp morphologies; and (3) an enhanced CLIP visual adapter with per-layer adaptive feature fusion and generalized mean pooling to capture multi-scale features for better polyp localization. During training, we use one annotated colorectal polyp dataset (e.g., CVC-ColonDB) to learn patch-level image-text correspondence. The model is then evaluated in a zero-shot manner on different polyp datasets (CVC-ClinicDB, Kvasir-SEG, and CVC-300), where we evaluate pixel-level anomaly detection performance.

The framework demonstrated robust zero-shot generalization on unseen test cohorts. Without any dataset-specific fine-tuning, the model achieved a mean pixel-level AUROC of 0.94 and a mean average precision of 0.81 across the 12 leave-one-dataset-out zero-shot transfer settings. In the CVC-ColonDB-source benchmark, the model achieved a mean Dice coefficient of 0.84 across CVC-ClinicDB, Kvasir-SEG, and CVC-300. This high level of performance was consistent across datasets with distinct visual characteristics, underscoring the ability of the model to detect diverse polyp morphologies that it had not been explicitly trained to recognize.

Our findings demonstrate that an anomaly-aware vision-language model significantly improves cross-dataset polyp detection generalization without requiring normal images for training. This multimodal strategy may facilitate the robust deployment of artificial intelligence-based colorectal screening systems by enabling reliable detection of diverse polyp morphologies across different clinical settings. Extension to non-polyp colorectal pathologies (e.g., ulcerative colitis and colorectal tumors) remains an important direction for future work, pending the availability of pixel-level annotated datasets for these lesion categories.

Generative artificial intelligence (AI), particularly large language models (LLMs) and multimodal systems, is emerging as a potentially important innovation in intensive care medicine. The intensive care unit (ICU) is a data-dense, high-acuity setting where rapid and accurate decisions are critical. These models can translate complex multimodal data into interpretable and clinically actionable insights across diagnostic, prognostic, and documentation workflows. This review outlines six key domains in which generative AI is currently being explored for its potential to reshape critical care: clinical decision support; clinical documentation automation (AI scribe, voice-to-note); predictive analytics, including sepsis and acute respiratory distress syndrome prediction, acute kidney injury management, ventilator liberation readiness, delirium monitoring, and continuous renal replacement therapy optimization; ICU data summarization and multimodal monitoring; synthetic data generation; and legal and ethical governance. In clinical decision support, hybrid models that integrate time-series monitoring data with LLMs can contextualize alerts, generate diagnostic suggestions, and offer treatment plans with explainable reasoning. Documentation tools that leverage ambient listening and voice-to-note AI can streamline progress notes and discharge summaries, thereby reducing clinician workload. In predictive analytics, LLMs enhance model performance by augmenting sparse electronic health record data and translating outputs into interpretable narratives. Synthetic data generation enables algorithm development and training, particularly for rare events, while protecting patient privacy. However, the realism and ethical deployment of such data require rigorous validation. Widespread implementation of generative AI will require careful attention to challenges related to trust, validation, bias, liability, and regulatory compliance. The use of these tools must remain under clinician supervision to ensure transparency and accountability. With responsible deployment, generative AI may augment ICU workflows, improve outcomes, and reduce clinician burden, potentially becoming an indispensable component of critical care delivery.