PURPOSE: Accurate preoperative risk stratification for post-hepatectomy liver failure (PHLF) is crucial. This study aimed to investigate whether, in patients deemed eligible for safe resection by conventional clinical criteria, ⁶⁸Ga-FAPI-04 PET/CT-based quantitative assessment of liver fibrosis could serve as a non-invasive predictor for PHLF. METHODS: In this retrospective study, 107 patients with liver cancer underwent preoperative ⁶⁸Ga-FAPI-04 PET/CT. Quantitative parameters, including liver SUVmax, SUVmean, SUVpeak, and liver-to-aorta SUV ratios (TBRs), were measured. These imaging metrics were correlated with histopathological fibrosis stage (S0-S4) and analyzed alongside clinical and surgical variables to identify independent predictors of PHLF. RESULTS: PHLF occurred in 30 (28.0%) patients. All PET parameters significantly increased with advancing fibrosis stage (all P < 0.05). Multivariable logistic regression identified TBRmax (odds ratio [OR] = 4.590, 95% CI 1.439 to 14.637, P = 0.010), prothrombin time (OR = 1.566, 95% CI 1.009 to 2.428, P = 0.045), and major hepatectomy (OR = 2.857, 95% CI 1.065 to 7.665, P = 0.037) as independent predictors of PHLF. A nomogram incorporating these three variables demonstrated good predictive performance, with an area under the curve of 0.780, satisfactory calibration, and positive net benefit on decision curve analysis. CONCLUSION: Preoperative ⁶⁸Ga-FAPI-04 PET/CT offers a histologically confirmed, non-invasive method to quantify liver fibrosis. This hypothesis-generating model offers a promising approach for preoperative risk stratification, warranting further validation in independent cohorts.

Accurately predicting long-term outcomes after stroke remains a key challenge in personalized medicine. Here, we present a neuroimaging platform that forecasts individualized cognitive outcomes in patients with ischemic stroke using deep learning-based lesion segmentation and location-/network-based features. This novel, fully automated system is capable of processing raw DICOM MRI data from heterogeneous scanners and generating text-based, personalized outcome information. To demonstrate this pipeline, we trained cognitive outcome-prediction models using a large lesion cohort (N = 604) and applied them to an independent stroke cohort (N = 153). Multiple cognitive outcome predictions achieved reasonable accuracy, with 96% concordance with manual methods. A report generated by a large language model provides interpretable, patient-specific prognoses within ~3 min. This demonstrates the potential for imaging-informed prognostication to inform stroke care and guide rehabilitation strategies.

Virtual simulation (VS) is pivotal in undergraduate dental education for cognitive and preclinical training. Expanding VS adoption and artificial intelligence (AI) incorporation necessitate an evidence synthesis to clarify current landscapes, challenges, and future directions. Following the PRISMA-ScR guidelines and Arksey and O'Malley's framework, this scoping review included 57 studies (2007-2025) to map the landscape of VS and its AI-driven future. The included VS systems could be broadly categorized into two generations, with first-generation systems supporting knowledge acquisition and second-generation systems supporting procedural skill development. VS applications spanned multiple dental subspecialties with uneven technological maturity and depth. With assessment of examinations, expert-rated performance, simulator-derived metrics, and learner-reported questionnaires, most studies reported that VS improved learner performance and confidence. Challenges persist regarding uncertain long-term knowledge retention, incomplete replication of clinical settings, resource-intensive implementation, and the limited role in education. This review further explored the implications of future AI-VS incorporation, including AI-driven virtual patients and tutors, fidelity enhancement, adaptive feedback, and personalized competency-based assessment.

PURPOSE: Data on the use of fluorodeoxyglucose (FDG) positron emission tomography (PET)/computed tomography (CT) in non-granulomatous myocarditis are limited, and no study has evaluated the diagnostic performance of combined FDG-PET and rest myocardial perfusion imaging (rMPI). METHODS: We conducted a multicenter prospective observational feasibility study of patients who underwent combined FDG-PET/CT with rMPI as well as cardiac magnetic resonance (CMR) for suspected myocarditis. When available, endomyocardial biopsy (EMB) established the diagnosis according to consensus criteria. Otherwise, the final adjudicated clinical diagnosis served as the reference standard. All patients had a minimum follow-up of 12 months following PET imaging. RESULTS: Twenty-five patients were included, of whom 18 (72%) were diagnosed with myocarditis. FDG-PET and rMPI demonstrated a sensitivity/specificity of 61%/100%, and 59%/71%, respectively. Inter-reader agreement for FDG-PET was excellent (κ = 0.84, 95% CI 0.62-1.00). The combination of FDG-PET and rMPI yielded a sensitivity of 82% (95% CI: 50-93%) and a specificity of 71% (95% CI: 29-96%). Late gadolinium enhancement (LGE), excluding isolated subendocardial patterns, showed a sensitivity/specificity of 72%/57%. Spatial agreement between rMPI perfusion defects and LGE on CMR was modest (κ = 0.35, 95%CI: 0.23-0.47). CONCLUSION: Combining FDG-PET with rMPI may be a feasible imaging modality for detecting non-granulomatous myocarditis by capturing both inflammatory and fibrotic components across different disease stages.

The discrepancy between serum triglyceride levels and the clinical severity of hyperlipidemic acute pancreatitis (HLAP) complicates risk stratification. Traditional lipidomics, which primarily rely on linear abundance, often fail to distinguish the HLAP-specific lipidome from the metabolic background of hypertriglyceridemia (HTG). To overcome this limitation, DeepLipiDecipher was developed as a knowledge-guided graph neural network framework that integrates lipid chemical structures with metabolic topology to identify latent lipotoxic features. In a retrospective cohort of 433 subjects, DeepLipiDecipher demonstrated robust classification performance (AUC = 0.810), effectively distinguishing the HLAP phenotype and outperforming conventional machine learning models. Interpretability analysis revealed that HLAP susceptibility correlated with a distinct structural lipid profile, marked by the synergistic enrichment of polyunsaturated and ether-linked phospholipids, rather than total lipid mass. Moreover, computational causal inference implicated a pathogenic mechanism wherein SMPD3-mediated ceramide accumulation induced basal cytotoxicity, and PTGS2 hyperactivation promoted the peroxidation of these vulnerable lipids, triggering systemic inflammation. These results highlight the value of incorporating network topology into lipidomic analysis and suggest the network-inferred SMPD3-Ceramide-PTGS2 immunometabolic axis as a potential therapeutic target for preventing the progression from metabolic dysfunction to acute organ injury.

OBJECTIVES: Large language models (LLMs) are increasingly explored as decision-support tools in medical imaging. However, their ability to align with country-specific guidelines, which often diverge, remains uncertain. We set out to evaluate the geographic neutrality of three state-of-the-art LLMs-GPT-o3, Mistral Large, and DeepSeek R1-and a biomedical LLM (MedGemma 1.5 4B), when applied to neuroradiology scenarios with conflicting U.S. and non-U.S. MATERIALS AND METHODS: Vignettes derived from contradictory international guidelines were presented to each model under two conditions: an implicit setting, where no guideline was specified and vignettes were provided in English and French; and an explicit setting, where prompts directed models to follow a named guideline. Performance was reviewed against the target guideline, and mitigation strategies were tested. RESULTS: Thirty clinical vignettes presenting conflicting guidelines were evaluated by GPT-o3, Mistral Large, and DeepSeek R1. In the implicit setting, all models favored U.S. guidelines, with GPT-o3, Mistral, and DeepSeek aligning with them in 27 of 30 scenarios (90.0%; 95% CI, 74.4-96.5). In the explicit setting, adherence declined sharply for non-U.S. recommendations for all models. Providing the complete guideline text was the most effective mitigation strategy, restoring accuracies above 90% across all models. CONCLUSION: Across languages and model origins, LLMs exhibited a systematic bias toward U.S. neuroradiology guidelines, even when explicitly instructed otherwise. This U.S.-centrism likely reflects training data imbalances and raises concerns for safe global deployment. Strategies for local contextualization, such as guideline integration at deployment, are necessary to ensure context-appropriate clinical decision support. KEY POINTS: Question Do large language models display geographical neutrality in neuroradiology decision support? Findings Even models developed in France and China systematically preferred United States guidelines, aligning with them in most implicit scenarios while failing to follow explicit guidelines from other sources. Clinical relevance This systematic United States-centric bias poses clinical and legal risks for global deployment. Safe implementation requires specific localization strategies, such as providing full guideline texts, to ensure recommendations align with local practice standards.