Background: Technical and regulatory constraints limit application of large language models (LLMs) for augmenting diagnostic reasoning in radiology. Reader-mediated text-based workflows may provide a practical alternative. Objective: To evaluate the impact on diagnostic performance of LLM assistance using reader-generated free-text image descriptions and to assess the effect of reader expertise on this LLM-augmented diagnostic workflow. Methods: This retrospective study included 93 cases (encompassing radiographic, CT, MRI, and PET/CT images) from the Korean Society of Thoracic Radiology quiz platform from January 2014 to December 2017. Five differential diagnoses (correct diagnosis and four distractors) were assembled for each case. Ten readers (five thoracic radiologists, five radiology residents) independently interpreted cases. In session 1, readers selected the most likely diagnosis and provided a free-text description of key findings. An LLM (Gemini 3.0 Pro) was inputted the free-text description-without case images-and outputted a ranking of the five differential diagnoses along with explanatory rationales for the top-three options. In session 2, readers were provided the LLM output from their own free-text description and re-selected a most likely diagnosis. LLM performance using images-without free-text descriptions-was also assessed. LLM accuracy was determined using top-ranked diagnoses. Reader groups were compared using generalized estimating equations. Results: LLM accuracy was 52.7% when inputted images and 63.9% when inputted reader-generated descriptions. LLM accuracy was greater using descriptions generated by thoracic radiologists than by residents (67.3% vs. 60.4%; P<.001). From session 1 to session 2, accuracy increased for thoracic radiologists from 56.3% to 65.6% and for residents from 42.4% to 58.5%, respectively. Accuracy improvement between sessions was greater for residents than thoracic radiologists (16.1 vs 9.2 percentage points; P=.02). Residents, compared with thoracic radiologists, demonstrated a greater rate of accepting LLM-favored diagnoses (73.2% vs 48.9%; P<.001), including a greater rate of switching to an incorrect diagnosis following misleading LLM output (60.6% vs 32.4%; P=.009). Conclusions: The text-based LLM-assisted workflow yielded improved reader accuracy although was heavily influenced by reader expertise. Clinical Impact: The utility of human-in-the-loop workflows arises from dynamic reader-LLM interactions shaped by the expertise of the operator formulating model inputs and critically evaluating model outputs.
Advances in MRI hardware and acceleration strategies have enabled substantial reductions in musculoskeletal MRI acquisition times over the past decade. Advanced acceleration techniques have facilitated four- to eightfold acceleration but are often limited by noise amplification and reconstruction artifacts at higher acceleration factors. The clinical introduction of deep learning (DL)-based image reconstruction addresses traditional constraints by improving SNRs, reducing artifacts, and enhancing image quality, thereby enabling higher acceleration factors than previously achievable with conventional reconstruction methods. DL reconstruction and superresolution techniques allow comprehensive musculoskeletal MRI protocols to be performed in less than 10 minutes across a range of applications, field strengths, and vendors. Successful implementation requires consideration of hardware capabilities, anatomic constraints, protocol design, and workflow adaptation to fully realize efficiency gains. In addition to technical factors, operational considerations-including scheduling logistics and infrastructure adjustments-are important to translate scan time reductions into clinical value. Early validation studies show preserved or improved diagnostic performance of DL-accelerated MRI compared with conventional protocols, supporting their growing integration into clinical practice. Continued technical development and clinical validation will further define the role of DL reconstruction and potentially facilitate even greater acceleration and efficiency gains.
Background: Neonates' immature renal function may cause increased susceptibility to renal toxicity from iodinated contrast media (ICM). However, little high-quality evidence addresses acute kidney injury (AKI) risk after neonatal ICM exposure. Objective: To evaluate associations between ICM administration during CT and AKI risk in ICU-admitted neonates using contemporary neonatal AKI criteria and propensity-score analysis to address potential confounding factors. Methods: This retrospective study included neonates admitted to the ICU who underwent contrast-enhanced CT, noncontrast CT, or noncontrast MRI between June 2000 and June 2023. Examinations were stratified into ICM and control groups. AKI was defined based on contemporary neonatal-specific criteria as an increase in serum creatinine of ≥0.3 mg/dL within 48 hours relative to the examination's immediately preceding value or of ≥50% within 7 days relative to the lowest preexamination value since birth. AKI was classified as stage 1, 2, or 3, reflecting increasing severity. Two additional cohorts were formed to reduce confounding based on patient and examination characteristics: a propensity-score matching (PSM) cohort that contained reduced subsets of matched examinations and an overlap-weighted cohort that retained all examinations albeit with weighting reflecting covariate overlap. Associations between ICM exposure and AKI were assessed using logistic regression analyses. Results: The analysis included 1708 neonates (median age, 5 days; 987 male, 721 female) who underwent 1809 examinations. The ICM group included 1381 contrast-enhanced CT examinations; the control group included 428 noncontrast CT (n=41) or noncontrast MRI (n=387) examinations. The PSM cohort included 284 examinations per group. AKI incidence was significantly higher in the ICM group than in the control group in the original cohort (13.9% vs 8.2%; OR=1.83 [95% CI: 1.24-2.68]; P=.002), PSM cohort (12.3% vs 7.0%; OR=1.86 [95% CI: 1.03-3.36]; P=.04) and overlap-weighted cohort (12.8% vs 7.6%; OR=1.74 [95% CI: 1.05-2.86]; P=.03). Incidence of stage 2-3 AKI did not differ significantly between the ICM and control groups in any cohort. Conclusion: For neonates in the ICU, ICM exposure during CT was associated with increased risk of AKI, although not with stage 2-3 AKI. Clinical Impact: The analysis provide robust controlled evidence that neonatal ICM exposure increases AKI risk.
Mesenteric ischemia (MI) represents a spectrum of vascular intestinal injuries for which delayed diagnosis remains the leading determinant of outcomes. Despite advances in multidetector CTA technology and endovascular therapy, substantial variation persists in how mesenteric ischemia is defined, imaged, interpreted, and incorporated into clinical pathways. This AJR Expert Panel Narrative Review provides radiologists with pragmatic guidance for mesenteric ischemia evaluation on imaging. Key issues considered include the pathophysiologic spectrum and diagnosis of acute mesenteric ischemia (AMI); CT protocol optimization (including the role of spectral CT); vascular pathologies that lead to occlusive or nonocclusive AMI; CT features associated with ischemia and necrosis; scoring systems for estimating necrosis likelihood; interventional and surgical management; evaluation of postischemic reperfusion; chronic mesenteric ischemia; and the emerging role of dedicated mesenteric stroke centers in reducing time to revascularization. A simplified four-phase framework is presented to illustrate the temporal progression of CT findings in occlusive arterial AMI from occlusion without visible bowel ischemia, to early ischemia, late ischemia, and irreversible transmural necrosis. The article concludes with consensus statements addressing image acquisition, interpretation, and reporting, with the aim of supporting radiologists who serve as first-line decision makers in the diagnosis and management of mesenteric ischemia.
BACKGROUND. Prior studies using habitat imaging for lung nodule characterization have been limited by intermixing of nodule types, insufficient handling of lesion heterogeneity, and incomplete consideration of high-risk histologic features. OBJECTIVE. To predict high-risk features within lung adenocarcinoma presenting as part-solid nodules on low-dose CT (LDCT) using habitat imaging incorporating attenuation and entropy measurements. METHODS. This retrospective study included 781 patients (median age, 58 years; 266 men, 515 women) with 781 resected adenocarcinomas manifesting as part-solid nodules on LDCT from July 2018 to December 2025. Patients from one center formed a training set (n=578) and from three other centers an external test set (n=203). The outcome was high-risk adenocarcinoma, defined as poorly differentiated invasive adenocarcinoma or invasive adenocarcinoma with visceral pleural invasion, spread through air spaces, lymphovascular invasion, or lymph node metastases. K-means clustering was used to determine optimal nodule subregions for maps of attenuation and entropy (reflecting local heterogeneity); attenuation and entropy subregions were integrated to form habitats. Habitat volumes and volume ratios (relative to whole-nodule volumes) were determined. A model incorporating demographic and conventional CT features was constructed by multivariable logistic regression analysis. AUCs were compared using DeLong test. RESULTS. The optimal number of clusters for both attenuation and entropy was 2 (attenuation threshold, -447 HU; entropy threshold, 4.201), yielding four habitats (high-attenuation high-entropy, high-attenuation low-entropy, low-attenuation high-entropy, low-attenuation low-entropy). In the external test set, AUC for high-risk adenocarcinoma was significantly greater (p<.05) for high-attenuation low-entropy habitat volume (0.863) than for conventional CT features (nodule diameter, solid-component diameter, consolidation-to-tumor ratio, whole-nodule volume) (0.693-0.809), other habitat features (0.614-0.842), and the conventional model (comprising sex, solid-component diameter, and whole-nodule volume; 0.810). High-attenuation low-entropy habitat volume had sensitivity, specificity, PPV, and NPV of 86.4%, 68.6%, 43.2%, and 94.8%, respectively, in the external test set. An executable software application for the final analytic pipeline and corresponding source code were made publicly available (https://github.com/mzi969/Habitat-Imaging-High-Risk-LUADs). CONCLUSION. The high-attenuation low-entropy habitat volume outperformed conventional CT features in predicting high-risk histologic characteristics of adenocarcinoma. CLINICAL IMPACT. Habitat imaging could inform noninvasive risk stratification and clinical decision-making for part-solid nodules encountered during lung cancer screening.
BI-RADS v2025 updates the established BI-RADS framework to reflect contemporary breast imaging practice across mammography, ultrasound, MRI, and contrast-enhanced mammography (CEM). This review summarizes the principal cross-modality and modality-specific changes introduced in the new Manual and discusses their implications for interpretation, reporting, multidisciplinary communications, and audits, with an emphasis on new descriptor terminology, assessment clarifications, and modality comparisons. Key cross-modality updates include structured clinical indication categories, revised and standardized report organization, harmonized terminology, refined morphologic descriptors, refined assessment categories (e.g., clarification of BI-RADS categories 0 and 6; introduction of BI-RADS 4 subclassification for breast MRI to mirror other modalities), structured lesion localization, tissue composition assessment, lymph node reporting, and expanded audit methodology. Modality-specific changes include refined mammographic characterization for digital breast tomosynthesis, revised calcification terminology, recognition of nonmass lesions and perilesional echogenic features for ultrasound, introduction of enhancement and T2-related descriptors for MRI, and formal incorporation of CEM into the BI-RADS reporting framework. Overall, BI-RADS v2025 preserves the core principles of prior editions of structured reporting, evidence-based assessment categories, and linkage between imaging findings and management recommendations while improving reporting consistency and reproducibility, cross-modality correlation, and auditability and performance monitoring across the full spectrum of breast imaging modalities.
How do we teach effectively without sacrificing efficiency or burning out? Scott Simpson, DO, MSEd, speaks with host Surbhi Raichandani, MD, about radiology readout frameworks, formative feedback, workflow-conscious teaching, and preparation of trainees for independent practice in modern radiology.
Radiology strongly impacts patient care. However, radiologists' potential to promote high-value care has long been underappreciated, partly related to difficulties in establishing their role in driving patient-oriented outcomes (e.g., reduced morbidity and mortality; improved quality of life). Recognizing this gap, the American College of Radiology convened the Relevance and Impact Committee to develop a framework for measuring radiologists' impact. The committee proposed two foundational concepts: diagnostic imaging provenance (the lineage underlying a diagnosis being established) and relevance (the manner in which a diagnosis relates to upstream reasons for examinations and downstream decision-making). The committee applied these concepts to develop the medical imaging life cycle framework as a patient and practitioner centric feedback loop whereby each examination serves as a potential care inflection point and pertinent positive and negative findings alter patient trajectories. Advances in data science, including large-scale EHR-driven analyses and generative artificial intelligence, now aid tracking of such relationships, helping to operationalize a longstanding vision of linking imaging to outcomes. This Perspective explores these core concepts and terminology underlying measurement of radiology's impact and seeks to provide a roadmap for their implementation within modern healthcare systems. Such application will promote radiologists' recognition as proximal agents of clinically impactful high-value care.
In 168 women from three centers with isolated axillary lymphadenopathy and without a known breast cancer diagnosis, axillary lymph node biopsy yielded malignancy in 13.1%. Malignancy showed significant independent associations with age of 50 years or older (OR = 9.85), effaced fatty hilum (OR = 3.47), and increased cortical vascularity (OR = 6.63).
Opportunistic screening leverages existing imaging examinations performed for unrelated routine clinical indications to systematically extract quantitative biomarkers. Artificial intelligence tools have made deployment at scale increasingly feasible. However, the pathway from a validated algorithm to a functioning clinical program remains poorly defined, and prospective implementation at scale is uncommon. Successful deployment requires coordinated engagement from radiologists, information technology and operational teams, and clinical care teams, each facing distinct decisions that determine whether a program functions reliably and delivers patient benefit. This article presents a practical framework for opportunistic screening implementation organized around these three stakeholder groups. We apply this framework to opportunistic CT osteoporosis screening, drawing on our experience developing such a program at a large academic medical center. The framework presented is intended to be broadly applicable across opportunistic screening applications as the field moves from algorithmic validation toward clinical translation.
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
In the large multicenter Cosmos EHR dataset, neuroimaging utilization per 1000 ED encounters increased from 175 examinations in 2016 to 316 examinations in 2025. Utilization also increased for all assessed examination categories (e.g., for noncontrast head CT from 86 to 129 and for head-and-neck CTA from 12 to 52).
暂无摘要（点击查看详情）
Since the first FDA approval of a PSMA-targeted PET radiopharmaceutical ([68Ga]Ga-PSMA-11) in 2020, PSMA PET/CT has rapidly become the standard of care for certain indications in prostate cancer imaging. Subsequently, the FDA also approved [18F]DCFPyl and [18F]rhPSMA-7.3. Although these small-molecule agents share an amino acid binding domain specific to the PSMA molecule, they differ in key ways (e.g., radioisotope physical and chemical properties, ligand structure), with potential clinical implications regarding synthesis, distribution, and imaging characteristics. For example, differences in radiopharmaceutical clearance (predominantly renal vs hepatic) may impact the detection of lesions adjacent to the urinary tract or within the liver. This variability also influences pre-therapy assessment, during which lesion uptake is often compared to the liver as a reference. Direct intrapatient comparisons among PSMA-targeted radiopharmaceuticals remain limited, and current evidence suggests lack of overall superiority in diagnostic performance metrics for any individual radiopharmaceutical. Consequently, radiologists are left wondering why multiple PSMA-targeted radiopharmaceuticals are being developed, whether their distinct properties matter clinically, and ultimately which one to use. In this article, we explore clinically relevant similarities and differences of PSMA-targeted PET radiopharmaceuticals, highlighting their imaging characteristics, supply logistics, and other operational considerations, to guide their successful integration into clinical care.
暂无摘要（点击查看详情）