This cross-sectional study compares 3 large language models on the accuracy of their responses to optical coherence tomography image–based and non–image-based questions.

This cross-sectional cohort study systematically assesses data availability and clinical heterogeneity across 2 major oropharyngeal cancer datasets to evaluate impact on survival and develop recommendations for multi-institutional research design.

The role of AI in writing and medicine is discussed by JAMA+ AI Editor in Chief Roy Perlis, MD, MSc, and Kevin Roose, a technology journalist formerly with The New York Times, for this episode of the JAMA+ AI Conversations podcast.

This cohort study evaluates the association of denosumab therapy duration with the risk for osteonecrosis of the jaw and skeletal events among men with bone-metastatic castration-resistant prostate cancer.

This Viewpoint issues a call for community-wide coordination to evaluate large language models for mental health care.

This prognostic study evaluates whether relapse within 2 years can be predicted among patients with major depressive disorder by using a comprehensive machine learning pipeline that integrates an unprecedented set of multimodal data.

This Viewpoint recommends policies that would redistribute the financial and oversight burdens of health care artificial intelligence (AI) to AI developers and federal agencies so that all health systems can fully realize its benefits.

This cross-sectional study examines clinician use of an artificial intelligence–enabled clinical information platform, including the distribution of use across the clinical workforce by clinician type and specialty and the purpose of submitted queries.

This cross-sectional study assesses perceptions on use of artificial intelligence (AI) for medical patient portal messaging and mental health support, with opinions stratified by birth generation.

This Viewpoint discusses the use of artificial intelligence (AI) in medical education, and how physician training should change in parallel with AI use including emphasizing skills that AI makes more valuable, cultivating those it makes newly necessary, and carefully de-emphasizing those it makes less important.

This Viewpoint describes potential dangers of mental health artificial intelligence agents and the need to develop regulations to address these dangers.

This Viewpoint addresses poor mental health outcomes among youths who use artificial intelligence (AI) chatbots and explores the AI safety failures that warrant evaluation and regulation to improve such outcomes.

This cohort study examines automated 3-dimensional soft tissue phenotyping as extracted from routine computed tomography scans and body composition and mortality risk for patients with head and neck cancer.

This randomized clinical trial assesses the effect of an electronic chatbot-based cancer risk communication tool on knowledge of cancer risk among adolescents and young adults (AYAs) with a hereditary cancer syndrome.

This cohort study estimates the prevalence of AI psychosis, characterizes affected patients, and defines the role of conversational AI in documented psychotic episodes.

This Medical News article discusses how children are using artificial intelligence (AI) chatbots for emotional support and the challenges of developing evidence and mental health safeguards quickly enough to keep pace with the technology.

This cohort study of patients with benign or low-grade pancreatic head or uncinate process tumors assesses short-term surgical outcomes and long-term quality of life after enucleation vs pancreatoduodenectomy.

Nature Machine Intelligence, Published online: 23 October 2026; doi:10.1038/s42256-026-01331-8

Ballester and Tran-Nguyen advocate for transparency in data, code and experimental procedures in the evaluation of protein–ligand predictions. They discuss good practices for preventing data leakage and bias, enabling fair, reproducible and meaningful comparisons across different scoring methods.

Nature Machine Intelligence, Published online: 23 October 2026; doi:10.1038/s42256-026-01326-5

Li et al. introduce EdenGNN-Uni, a universal charge density model for the periodic table that is over an order of magnitude faster than some existing methods, facilitating efficient and accurate electronic structure calculations of materials.

Nature Machine Intelligence, Published online: 20 October 2026; doi:10.1038/s42256-026-01319-4

A neural network approach bridging Lagrangian and Eulerian perspectives transforms sparse observations into physically accurate, high-resolution maps of complex and chaotic fluid motions.

Nature Machine Intelligence, Published online: 20 October 2026; doi:10.1038/s42256-026-01318-5

Sharp et al. investigate the implications of unequal access to AI agents and of differences in their quality and number, developing a conceptual framework to identify potential sources of inequality and suggesting a research agenda to help address the resulting societal challenge.

Nature Machine Intelligence, Published online: 20 October 2026; doi:10.1038/s42256-026-01316-7

Wan et al. introduce a physics-informed learning model that simultaneously predicts long-range particle paths and flow fields based on very limited trajectory measurements, validated in both simulations and blood-flow experiments.

Nature Machine Intelligence, Published online: 19 October 2026; doi:10.1038/s42256-026-01333-6

Frontier LLMs are trained to solve problems in a step-by-step way, generating reasoning text in human-understandable language. However, these reasoning tokens are not accurate explanations of LLMs’ internal computations, raising challenges for monitorability.

Nature Machine Intelligence, Published online: 19 October 2026; doi:10.1038/s42256-026-01311-y

This work studies the reusability of CORNETO, a previously developed framework for learning biological networks from multi-omics data. Robust network inference is found to depend on data availability, prior knowledge, regularization and biological context.

Nature Machine Intelligence, Published online: 16 October 2026; doi:10.1038/s42256-026-01325-6

Li et al. introduce a neural-hierarchical skin that combines high-density iontronic sensor array, fast readout and data compression, enabling humanoid robots to perceive fine textures, manipulate delicate objects and respond to nociceptive stimuli.

npj Digital Medicine, Published online: 23 October 2026; doi:10.1038/s41746-026-03408-5

Human oversight of erroneous clinical AI and decision support a systematic review and meta analysis

npj Digital Medicine, Published online: 23 October 2026; doi:10.1038/s41746-026-03342-6

Technology-based structural racism and telehealth video usage among primary care patients

npj Digital Medicine, Published online: 23 October 2026; doi:10.1038/s41746-026-03409-4

Clinical AI governance for scalable innovation in value based healthcare systems

npj Digital Medicine, Published online: 23 October 2026; doi:10.1038/s41746-026-03372-0

From simulation to pedagogy: multi-model and randomized evaluation of structured AI standardized patients

npj Digital Medicine, Published online: 23 October 2026; doi:10.1038/s41746-026-03403-w

A deep learning framework for automated cervical spinal cord vertebral labeling and morphometric extraction from brain MRI with large-scale multicenter validation

npj Digital Medicine, Published online: 22 October 2026; doi:10.1038/s41746-026-03349-z

Large language models for multimodal clinical predictive modeling: a systematic review

npj Digital Medicine, Published online: 22 October 2026; doi:10.1038/s41746-026-03371-1

Multi-scale integration identifies an ox-LDL-autophagy-inflammasome framework linking osteoporosis and neurodegeneration

npj Digital Medicine, Published online: 22 October 2026; doi:10.1038/s41746-026-03221-0

A hybrid framework for Turner Syndrome facial detection: integrating deep feature learning and geometric prior knowledge under data scarcity

Nature Medicine, Published online: 24 October 2026; doi:10.1038/s41591-026-04756-1

As presented at the ESMO Congress 2026: In a 5 year follow-up of survival and correlative data from the pivotal phase 3 RELATIVITY-047 trial comparing relatlimab plus nivolumab versus nivolumab alone in patients with metastatic or unresectable melanoma, statistically significant differences in progression-free survival, overall survival and melanoma-specific survival were seen with high baseline levels of LAG3 high CD8+ T cells associated with improved response in the combination arm.

Nature Medicine, Published online: 24 October 2026; doi:10.1038/s41591-026-04740-9

As presented at the ESMO Congress 2026: In this exploratory, nonrandomized phase 2a trial, treatment of patients with immune checkpoint inhibitor-naive muscle-invasive bladder cancer with the GDF-15-neutralizing antibody visugromab and nivolumab produced encouraging pathological complete response and radiographic response rates with no new safety signals, supporting further evaluation in larger randomized studies.

Nature Medicine, Published online: 23 October 2026; doi:10.1038/d41591-026-00054-y

A patient-facing conversational AI showed promising feasibility and safety in a single-center primary care setting, providing a basis for larger-scale evaluation.

Nature Medicine, Published online: 23 October 2026; doi:10.1038/s41591-026-04730-x

The human microbiome has emerged as one of the most consequential frontiers in biomedical science. An international group of leading microbiome scientists, convened by the US Department of Health and Human Services, presents a strategic vision for realizing its potential to transform public health.

Nature Medicine, Published online: 23 October 2026; doi:10.1038/s41591-026-04734-7

As presented at the ESMO Congress 2026, in a randomized phase 2 trial evaluating an off-the-shelf lymph node-targeting peptide vaccine in patients with mutated KRAS pancreatic cancer within 6 months of standard curative intent therapy, the primary endpoint of disease-free survival was not met, but prespecified subgroup analyses showed significant disease-free survival benefit in the vaccine arm for patients with circulating tumor DNA-negative disease and in those without prior radiation treatment.

Nature Medicine, Published online: 23 October 2026; doi:10.1038/s41591-026-04711-0

As presented at the European Society for Medical Oncology Congress 2026, in a phase 1/2 trial in participants with solid tumors, invika, a first-in-class bispecific agonistic antibody that selectively targets the prevalent Vβ6/Vβ10 T cell subsets, was well tolerated, enabling the selection of a recommended phase 2 dose for further testing, and showed a preliminary efficacy signal in both immunologically hot and cold tumors.

Nature Medicine, Published online: 23 October 2026; doi:10.1038/s41591-026-04707-w

A large comprehensive analysis of clinical trial samples from patients with non-Hodgkin lymphoma who were treated with anti-CD19 CAR T therapy shows that a high proportion of naive-like and stem cell memory-like CCR7+CD45RA+ T cells in the pre-infusion CAR T product correlates with improved and prolonged clinical benefit, which was validated in an independent external cohort including different products.

Nature Medicine, Published online: 22 October 2026; doi:10.1038/d41591-026-00053-z

Researchers identify biomarkers that explain how maternal health is linked to mammary gland function, breast milk composition and infant growth, highlighting maternal nutrition as a public health priority.

I trained as a physician before I built a company, and medicine taught me to think of the prescribing decision as something that happens in the room. Write the script, and the decision is made.  Practice taught me otherwise. Plenty of prescriptions I wrote were never filled. The patient had already been forming a view ...;

Picture this: a patient receives a text reminding them to arrive early for an imaging appointment. The message doesn’t say that they need a new authorization, where to check in, or whether they can take their morning medication. The patient calls the health system, waits on hold, repeats their information, and eventually reaches an employee ...;

Safety is foundational to effective healthcare delivery. Organizations must protect patients, provide staff with immediate access to help during workplace violence situations, behavioral escalations, or medical emergencies, and ensure critical equipment is available when and where it’s needed. Meeting these demands depends on timely, actionable insights that enable healthcare teams to make informed decisions, respond ...;

Introduction: The Shift Toward Data-Driven Orthopedics  The healthcare landscape is undergoing a fundamental transformation, shifting away from standardized treatment protocols toward highly individualized, patient-specific care models. In orthopedics and spinal medicine, this evolution is largely driven by rapid advancements in machine learning, quantitative data analytics, and computational imaging. Where traditional diagnostics relied heavily on static ...;

Artificial intelligence has moved from science fiction to an increasingly familiar part of everyday life. A few decades ago, the idea that computers could understand language, analyze enormous amounts of information, or make predictions based on patterns would have seemed almost impossible. Today, those capabilities are becoming part of how we work, communicate, and deliver ...;

Two numbers tell you where healthcare AI actually stands.  $995 million went into healthcare AI in 2025. Roughly 70% of it never reached scale.  That is not a technology failure. The models are good. The demos are impressive. The pilots start fast.  Then they stall, and 18 months later nobody can explain what changed for ...;

Most digital products are built around a familiar set of goals: reduce friction, keep people engaged and make every interaction feel easy. If users hesitate, leave or report discomfort, the system treats it as a problem to solve. That logic makes sense for online shopping or entertainment. In healthcare, it can lead us in the ...;

How AI Supports Care Coordination starts with one practical idea: let technology handle repetitive coordination work while people remain responsible for clinical judgment and patient decisions. AI can help teams organize information, spot gaps, and communicate faster without removing the human oversight healthcare requires. Where can AI help with everyday care coordination? AI can reduce ...;

For a decade, “AI in healthcare” meant a diagnostic algorithm reading a scan. In 2026, the technology that’s actually reaching the exam room is far less glamorous: software that listens to the doctor-patient visit and writes up the notes. Ambient clinical documentation — AI that captures a patient-clinician exchange, structures it into a chart entry, ...;

Objectives;This study addresses the issue of spatial misalignment between fractional anisotropy (FA) maps and tractography atlases in neuroimaging. We propose a CycleGAN-based method to generate FA and directionally encoded color (DEC) maps directly from T1-weighted MRI scans, applicable to both healthy and tumor-affected cases.Methods;A CycleGAN model was developed and trained using unpaired data to convert T1-weighted MRI images into FA and DEC maps. The quality of the generated maps was assessed using Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) metrics. Special emphasis was placed on evaluating performance in tumor regions. Radiological assessments were also conducted to determine the clinical applicability of the generated maps.Results;The CycleGAN-based model effectively generated high-fidelity FA and DEC maps, with robust SSIM and PSNR values, particularly in tumor regions. Radiological evaluations highlighted the potential of this approach to enhance clinical workflows, offering a reliable AI-driven alternative to traditional image processing methods.Conclusions;This CycleGAN-based approach successfully mitigates the problem of spatial misalignment by generating accurate FA and DEC maps from T1-weighted MRI scans. The method offers a promising pathway for integrating assessments of white matter integrity into predictive models without requiring additional imaging.Advances in knowledge;This study introduces the first use of CycleGAN for generating FA and DEC maps from T1-weighted images for tumor-affected cases, presenting a method that could enhance neuroimaging procedures by reducing the need for additional scans and improving efficiency in clinical settings.;

<span class="paragraphSection">Abstract;Radiation oncology has proven an early adopter of machine learning (ML) and deep learning (DL) owing to its image-centered workflow and data richness. DL auto-segmentation of normal structures is already guideline-endorsed, with self-configuring nnU-Net delivering near-state-of-the-art accuracy and widely adopted. In this review, we explore advances in ML algorithms in processing and interpreting imaging-based data beyond normal tissue auto-segmentation with applications upstream and downstream of this task, highlighting training data volume and best-in-class architectures. Applications discussed range from diagnostics with “virtual biopsies” and automated staging to radiotherapy planning with gross tumor volume and clinical target volume (CTV) prediction. We illustrate the developments of cutting-edge transformer-based designs and foundation models, but also the important role of lightweight, task-specific classical ML models. Data scarcity may be eased by developments in few-shot learning (FSL) and transfer learning strategies. We also consider how end-to-end automated CTV delineation may require an agentic approach to mimic the reasoning of a radiation oncologist, capable of processing multi-modal data with clinical context. Together, these advancements point toward a future where ML will become increasingly central to radiation oncology practice. However, implementation of these tools beyond normal tissue auto-segmentation remains limited and will require careful assessment and validation in the clinic.;

Objectives;To evaluate feasibility of a federated AI monitoring service (FAMOS) for post-deployment surveillance of third-party AI applications used in chest X-ray (CXR) interpretation.Methods;FAMOS was deployed at 2 NHS Trusts using a federated architecture enabling local data processing while maintaining data governance compliance. De-identified CXRs from patients aged >18 years were retrospectively identified, along with relevant patient attributes (age, sex, inpatient status, image orientation, season, artefact). Chest X-rays were processed by 2 AI applications to simulate real-world deployment. FAMOS analyzed input data, AI inference values, and longitudinal human-AI agreement as proxy indicators of data, prediction, and behavioral drift. Input monitoring used image feature embeddings analyzed with principal component analysis and Hotelling’s <span style="font-style:italic;">T;² statistics, while risk-adjusted cumulative sum charts were applied to sequentially monitor AI inference values. human-AI agreement was evaluated at 10 time points over 3 months.Results;Input monitoring identified a small proportion of outlier examinations, predominantly associated with modifiable image quality issues. Artificial intelligence inference values remained stable across most findings for both vendors, with limited drift events detected. Human-AI agreement patterns differed between sites, remaining stable at 1 site while increasing over time at another, suggesting evolving automation bias.Conclusions;Real-time, federated monitoring of deployed radiology AI systems is technically and operationally feasible within clinical environments. Multi-domain platform-based monitoring provides scalable, independent oversight and may function as an early-warning system supporting identification of emerging risks following deployment.Advances in knowledge;This study introduces a federated, multi-domain monitoring framework integrating sociotechnical indicators for continuous post-deployment surveillance of clinical radiology AI tools.;

Objectives;To calculate the distributions of leg bone lengths (as measured from dual-energy X-ray absorptiometry [DXA] images) and their relationship with height. Estimating human height from long bone measurements is widely used in forensic and biological anthropology.Methods;An automated system was developed to find the outlines of the femur and tibia in DXA images from the UK Biobank. These were used to measure the length of the bones. Images with implants, severe abnormalities, or missing data were excluded. Relationships between lengths and with height were estimated using linear regression.Results;Of the 73 405 available DXA images, we successfully measured both femurs and tibia in 66 730 cases (91%). The lengths of femurs and tibias for males and females each show normal distributions, with left and right being very similar. The mean absolute difference between measurements at 2 time points was below 1.8 mm (femur), 2.3 mm (tibia), showing the system is repeatable. Linear regression gives the relationships between individual bone lengths and between them and height.Conclusions;The developed software system and the resulting analysis provide promising insights into bone geometry using DXA imaging, enabling large-scale, precise bone measurements. These models are relevant for predicting height, a subject of interest in archaeology and forensic science.Advances in knowledge;This is the largest study of the distributions of leg bone lengths and their relationship to height that we are aware of.;

Purpose;To develop an approach for evaluating and reporting the quality of synthetic medical data (SMDs).Method;As SMDs become increasingly common, their responsible utilization benefits from a detailed characterization of their quality and value within the intended usage context. Here, we introduce an approach to assess SMDs across 7 dimensions: Congruence, Coverage, Constraint, Consistency, Comprehension, Compliance, and Completeness. We also describe an approach for reporting the quality of synthetic data to stakeholders.Results;We applied the proposed approach to 7 digital mammography datasets. The analysis underscored the strengths and limitations of each dataset across these dimensions. For example, while datasets generated by generative AI methods show high Congruence, they often fall short in Constraint and Completeness compared to those generated by knowledge-based methods. Additionally, our findings highlighted that the quality of generated data varies across different subgroups (eg, breast density) with certain subgroups (ie, dense and hetero) showing lower quality. This suggests that subgroup-specific fine-tuning of the generative process may have an effect on downstream tasks.Conclusions;The proposed approach provides stakeholders with a tool for assessing data quality, which may be used to better understand and analyse synthetic datasets..Advances in knowledge;This study introduces an approach for assessing SMD quality. The approach offers a practical mechanism for reporting SMD quality, which can enhance the reliability and usefulness of synthetic datasets.;

Objectives;The interpretation of contrast-enhanced mammography (CEM) images heavily depends on radiologists’ expertise, highlighting the need for automated tools to assist clinical decision-making. This pilot study aimed to develop a deep learning model using CEM images to predict the histological diagnosis of breast cancer.Methods;We retrospectively analyzed patients who underwent CEM followed by histopathological assessment (October 2022-May 2023) across 2 centers. CEM images from center 1 were used for model development, including training and 10-fold cross-validation, while images from center 2 served as an independent external test set. The breast region was manually segmented, and 2 deep learning architectures were implemented as ensemble classifiers, using biopsy results as reference standard. Performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the Receiver Operating Characteristic Curve (ROC) curve (AUC).Results;A total of 106 CEM images were retrospectively analyzed (50 from center 1 and 56 from center 2), obtained using different mammography systems. Histopathology identified 51 lesions (48%) as malignant and 55 (52%) as benign. Cross-validation yielded an ROC-AUC of 75% [58.4-91.6], accuracy of 68.7% [58.3-79], sensitivity of 68.9% [52.2-85.6], specificity of 67.8% [65.4-70.2], PPV of 69.3% [55.5-83.1], and NPV of 73.5% [55.2-91.8] (<span style="font-style:italic;">P; < .05). External testing on images from center 2 (27 malignant, 29 benign) achieved an accuracy of 64.3%.Conclusions;The study showed promising performance but requires further development and dataset expansion. The model has the potential to be integrated into clinical practice.Advances in knowledge;A deep learning model based on CEM provides a tool to support physicians in the decision-making process and reduce invasiveness in breast cancer management.;

Objective;This study developed a radiomics-based pipeline to identify suspicious findings on 2D screening mammograms by training models to distinguish radiologist-annotated abnormalities from normal breast tissue.Methods;A total of 1604 screening mammograms (<span style="font-style:italic;">n; = 1294 participants) were used in this retrospective study. Each mammogram included an original capture image, and a secondary capture image with a single radiologist-drawn annotation indicating a region of interest (ROI) with an abnormality. The annotation on each secondary capture image was used to select an ROI in the original image. An ROI with normal tissue was automatically selected from the remaining breast tissue for comparison. Radiomics features were extracted from the ROIs with abnormalities and normal tissue. Feature selection was performed using the SciKit-Learn SelectKBest method with Chi-squared, analysis of variance (ANOVA) <span style="font-style:italic;">F;-, and mutual information score functions. Logistic regression, random forest, XGBoost, bagging, discriminant analysis (DA), and support vector machine classifiers were trained on the selected features. The model performance was evaluated with the area under the receiver operating characteristic curve (AUC) on a holdout test set.Results;While AUC values ranged from 0.69 to 0.73, no significant differences were observed between models (DeLong test, <span style="font-style:italic;">P; > .05). The nominally highest performance was achieved via ANOVA <span style="font-style:italic;">F;-score feature selection and DA (AUC: 0.73; 95% CI, 0.70-0.77).Conclusion;The radiomics-based pipeline shows promise in distinguishing abnormalities from normal tissue on screening mammograms.Advances in knowledge;Radiomics has the potential to enhance breast cancer detection and is a step towards the integration of advanced machine learning into screening workflows.;

<span class="paragraphSection">Abstract;Artificial intelligence (AI) can revolutionize clinical workflows in radiology but requires organizational change. An institutional strategy to develop and evaluate AI tools is outlined. A multidisciplinary AI board with a patient and public involvement and engagement group was created. A comprehensive framework was formed, comprising workstreams covering information governance; technical rigor, performance, and safety; economic considerations; and ethical and medical-legal aspects. In addition to recurring meetings, a workshop with clinicians, information technology specialists, and patient representatives helped to identify priority use cases. Technical infrastructure was enhanced to support the development, performance assessment, and deployment of AI tools. Primary areas for AI applications included training staff, vetting of image requests, quality assurance, image interpretation, and communicating imaging findings to patients. Potential barriers, gaps in evidence, and subsequent actions for AI implementation were outlined. Avenues for collaboration with industry and market-available solutions were outlined. A virtual Picture Archiving and Communication System server was developed and then connected to a deployment platform for performance evaluation of AI products. Establishing an institutional AI board and imaging AI sandbox has guided safe, effective AI implementation while creating an ideal setting for innovation and industry partnership. Our approach to the integration of imaging AI provides a pragmatic guide for other institutions.;

Objectives;This study aimed to restore missing regions from the limited field of view (FOV) using image- and sinogram-based conditional GAN (cGAN) models.Methods;cGANs are deep learning frameworks that generate realistic data via a competitive neural network process. We used planning CT (pCT) datasets from 96 patients: 64 for training, 16 for validation, and 16 for internal testing. Two cGAN models (image-based and sinogram-based) were developed to generate body contour outside the FOV. Next, 23 cone-beam CT (CBCT) datasets were evaluated as an external test group.Results;In pCT internal test datasets, the median values for mean absolute error (MAE), root mean square error (RMSE), and structural similarity index measure (SSIM) for each model were as follows: image-based model—101.73 HU for MAE, 39.26 HU for RMSE, and 0.83 for SSIM; sinogram-based model—16.91 HU for MAE, 23.19 HU for RMSE, and 0.91 for SSIM. In CBCT external test datasets, the sinogram-based model outperformed the image-based model with a median MAE of 73.32 HU versus 180.72 HU, a median RMSE of 37.02 HU versus 43.42 HU, and a median SSIM of 0.75 versus 0.63. The sinogram-based model demonstrated significant improvements in MAE, RMSE, and SSIM (<span style="font-style:italic;">P ;< .05).Conclusions;The sinogram-based cGAN model exhibits considerable potential for restoring missing regions outside the FOV, outperforming the image-based model in accuracy metrics.Advances in knowledge;This model offers a novel approach to accurately predict missing regions from a limited FOV, enhancing continuity of the body contour while accommodating patient-specific variations.;

<span class="paragraphSection">Abstract;Chest X-rays (CXRs) are the most commonly performed imaging investigation. In the UK, many centers experience reporting delays due to radiologist workforce shortages. Artificial intelligence (AI) tools capable of distinguishing “normal” from “abnormal” CXRs have emerged as a potential solution. If “normal” CXRs could be safely identified and reported without human input, a substantial portion of radiology workload could be reduced.This article examines the feasibility and implications of autonomous AI reporting of “normal” CXRs, using the United Kingdom as an example setting. Key issues include defining “normal,” ensuring generalizability across populations, and managing the sensitivity-specificity trade-off. It also addresses legal and regulatory challenges, such as compliance with IR(ME)R and GDPR, and the lack of accountability frameworks for errors. Further considerations include the impact on radiologists practice, the need for robust post-market surveillance, and incorporation of patient perspectives. While the benefits are clear, adoption must be cautious, with strong governance, legal clarity, and rigorous clinical validation to ensure safe and sustainable use.;

Objectives;Artificial intelligence (AI) has shown promise for estimating volumetric breast density values from processed, “for presentation,” mammograms. However, previous evaluations have typically used small datasets or focused on a single vendor. In this study, we aimed to improve volumetric breast density estimation from processed mammograms for the three main UK vendors with a combination of improved training methods and the utilization of up-to-date data from the large OPTIMAM Mammography Image Database (OMI-DB).Methods;Paired processed/unprocessed mammograms were obtained from OMI-DB. Ground-truth, image-level density values were calculated by passing unprocessed images through a commercial density estimation tool. AI tools, comprising feed-forward convolution neural networks, were then trained to reproduce these values from the corresponding processed mammograms.Results;Patient-level AI predictions for volumetric breast density demonstrated strong correlation with ground-truth values derived from unprocessed image counterparts (<span style="font-style:italic;">r; = 0.954-0.976). Models trained on less prevalent manufacturers performed worse (<span style="font-style:italic;">r; = 0.954 compared to 0.976 for the most prevalent manufacturer), highlighting the importance of collecting larger training datasets in future. Error levels were higher in patients with dense breasts. Model performance was generally consistent across screening sites but correlated with patient age, possibly due to the correlation of age and breast density.Conclusions;The presented models demonstrated good performance overall and were generally consistent across screening sites.Advances in knowledge;The presented AI tools provide a means of estimating breast density from processed mammograms, enabling further research into breast cancer epidemiology and risk where only processed mammograms are available.;

<span class="paragraphSection">Abstract;This paper presents a reproducible, data-driven approach for prioritisation of AI-detected chest X-ray (CXR) findings to support faster lung cancer diagnosis in the NHS. The Annalise Enterprise CXR system was deployed in shadow mode across seven acute trusts in Greater Manchester. Two cohorts were used: a retrospective cancer cohort (<span style="font-style:italic;">n; = 1,282) with confirmed lung cancer and visible CXR abnormalities, and a prospective cohort (<span style="font-style:italic;">n; = 13,802) comprising consecutively acquired GP-referred CXRs. Prevalence ratios were calculated for 124 AI-detected abnormalities across both cohorts, and three prioritisation strategies were developed. Strategy 3, which combined prevalence analysis with expert clinical review, achieved optimal performance with a sensitivity of 95.87% and estimated specificity of 79.11%, while maintaining a negative predictive value of 99.95%, for identification of lung cancer. Findings most associated with cancer included solitary lung mass, mediastinal mass, and hilar lymphadenopathy. An Excel-based tool was developed to support rapid configuration and evaluation of categorisation. Application of this approach enabled safe deployment of AI using shadow mode to inform configuration prior to live use. This work provides a scalable model for AI implementation in radiology workflows that aligns with the National Optimal Lung Cancer Pathway and addresses real-world challenges of diagnostic capacity, safety, and reproducibility.;

Objective;To assess the cost-effectiveness of using artificial intelligence (AI)–derived software to assist reading CT scans of the chest to identify and analyse lung nodules compared to unaided reading in symptomatic, incidental and screening populations.Methods;Decision tree structures were developed in TreeAge Pro 2021. Structures were informed by British Thoracic Society clinical guidelines and clinical opinion. Results were presented as incremental cost-effectiveness ratios (ICERs) expressed as cost per quality-adjusted life-year (QALY) over a lifetime from the UK National Health Service and Personal Social Services perspective.Results;For the symptomatic population, the unaided radiologist reading strategy dominated the AI-assisted reading strategy. In the incidental population, unaided radiologist reading was cost-effective with an ICER of approximately £1000 per QALY. Conversely, in the screening population, AI-assisted radiologist reading dominated unaided reading. The cause of AI assistance being cost-effective depended on the number of people who had undergone CT surveillance because of non-cancerous findings. Given the limitations in the quality and quantity of evidence to inform inputs, these results should be interpreted with caution.Conclusion;Current analyses based on limited evidence suggested that, in the symptomatic and incidental populations, unaided radiologist reading may be the more cost-effective strategy, while in the screening population, AI-assisted radiologist reading appeared to be the dominant strategy. Better quality evidence is required to have a definitive answer about their cost-effectiveness.Advances in knowledge;This paper shows whether adding AI-derived software to radiologists' reading of CT scans to identify lung nodules offers good value for money.;

<span class="paragraphSection">Abstract;Radiology is undergoing a major shift with the growing use of artificial intelligence (AI), and more change is expected with the emergence of agentic AI—systems that can initiate, manage, and coordinate tasks. So far, most discussions about AI’s impact on radiology follow 2 main approaches. The first, the “displacement” approach, tries to predict which jobs are most at risk of being replaced by AI. This narrative often warns that radiologists may be displaced. The second, the automation-versus-augmentation approach, looks within jobs to identify which tasks are likely to be fully automated (automation) and which will be improved by AI working alongside humans (augmentation). This paper introduces a third approach: <strong>reconfiguration</strong>. Instead of focusing on job loss or task replacement, the reconfiguration model looks at how AI changes the way tasks connect, how responsibilities shift, and how professional roles evolve. Drawing on recent research and developments in AI, this paper advances the reconfiguration approach and articulates why it offers a clearer way to understand—and help shape—the future of work in radiology. This paper offers a forward-looking reflection on the shifting nature of radiological work—clinically, educationally, and organizationally—as AI systems become increasingly integrated into practice.;

Objectives;Deep learning neural network (DLNN)-based tools can automate body composition analysis for cancer cachexia research. We aimed to evaluate a DLNN tool trained on a European population of Chinese cancer patients.Methods;Computed tomography (CT) images at the 3rd lumbar vertebral (L3) level of Chinese gastric cancer patients were retrospectively collected. An externally validated DLNN tool (Mosamatic) was used to segment skeletal muscle, visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT). Manual segmentation was performed using SliceOmatic software (TomoVision, version 5.0). Geometric similarity between automated and manual segmentation, and the reliability was assessed.Results;The cohort comprised 203 patients with a median body mass index (BMI) of 22.2 kg/m², and 604 CT images at L3 were collected. The median Dice Similarity Coefficient (IQR) of skeletal muscle, VAT and SAT were 0.973 (0.961-0.980), 0.980 (0.964-0.989), and 0.967 (0.945-0.977), respectively. The median Lin’s Concordance Correlation Coefficient for skeletal muscle area (0.983), VAT area (1.000), SAT area (0.998), skeletal muscle radiation attenuation (0.995), VAT radiation attenuation (0.994), and SAT radiation attenuation (0.997) demonstrated excellent reliability. Low BMI (<18.5 kg/m²) and ascites impaired the agreement between the 2 methods. The automated method showed high diagnostic concordance with manual segmentation for sarcopenia (<span style="font-style:italic;">κ ;= 0.843, <span style="font-style:italic;">P ;< .001) and myosteatosis (<span style="font-style:italic;">κ ;= 0.946, <span style="font-style:italic;">P ;< .001).Conclusions;The Mosamatic tool displays excellent generalizability to analyse body compositions in Chinese gastric cancer patients and can facilitate cachexia research.Advances in knowledge;The Mosamatic tool displayed excellent generalizability without recalibration to analyse body composition on the 3rd lumbar vertebral CT images in Chinese gastric cancer patients.;

<span class="paragraphSection">Abstract;Recent advances in artificial intelligence (AI) offer significant potential to address the growing bottleneck in radiology caused by an increasing volume of imaging studies amidst a global shortage of radiology professionals. This study presents a comprehensive review of the latest developments in AI, particularly in vision-language models for radiology report generation, providing radiologists with a current reference. We conducted a focused literature search for studies published from 2020 to 2024 and included 14 studies in our review specifically on chest X-ray datasets with limited coverage of 3D modalities, reflecting the early stage of research and ongoing methodological advances in report generation for volumetric imaging. We analysed the model architectures, report generation capabilities, training datasets, evaluation metrics, and performance of these models. Our review highlights the evolution of AI in radiology report generation and underscores the critical need for diverse datasets and standardized evaluation metrics. Despite rapid progress, current AI models are not yet capable of consistently producing high-quality reports and require further improvements in data diversity, model training, and evaluation metrics to achieve a level comparable to human experts.;

Objectives;There is a growing need to develop user-friendly, bladder-specific image analysis tools that can produce reliable artificial intelligence (AI)-quantitative imaging biomarkers (QIBs) derived from multiparametric (mp)MRI data for clinical applications. To address it, we developed an AI-powered BLADdEr multiparametric MRI Analysis for Clinical Application (AI-BLADE, current release v1.0) toolbox designed for extracting mpMRI-derived quantitative metrics.Methods;AI-BLADE is an advanced tool for bladder-specific mpMRI data analysis with 2 core functionalities: (1) Deep Feature Analysis (MRI-DFA toolkit) and (2) Data-Driven Model-Based Analysis (MRI-MBA toolkit). AI-BLADE offers customizable options and serves as a one-stop shop solution for bladder cancer (BCa) clinical applications. The models within DFA and MBA were tested separately on 2 patient cohorts. DFA was used to classify BCa histology subtypes (<span style="font-style:italic;">n; = 104) with T2-weighted images, while MBA was used to interrogate tumour physiology by deriving mpMRI QIBs, including apparent diffusion coefficient (ADC), and volume transfer constant (K^trans) obtained from 34 BCa patients.Results;Out of the 17 AI models tested, the VGG19 model with a decision tree classifier and no feature selection for the fully connected layer 7 achieved the highest area under the curve of the receiver operating characteristic of 0.79 in classifying BCa histology subtypes, demonstrating the strongest performance. The mean ADC and K^trans values were 1.22 × 10⁻³ (mm²/s) and 0.27 (min⁻¹), respectively, reflecting underlying tumour physiology.Conclusion;The AI-BLADE (v1.0), a flexible and user-friendly software toolbox for analysing mpMRI data, shows strong potential for application in BCa oncology, offering capabilities that can enhance diagnostic accuracy and support improved patient outcomes.Advances in knowledge;This is the first study to design, develop, and implement a novel bladder-specific AI toolbox for analysing mpMRI data. AI-BLADE enables an advanced image analysis workflow, facilitating AI-QIB-based clinical decision-making for patients with BCa.;

Objectives;Deep learning models developed for the classification of radiological reports have lacked explainability. We aimed to validate and explain a pretrained classification model by applying it to the removal of confounding data from a radiological dataset.Methods;Two radiologists categorized 2038 anonymized MRI head free-text radiology reports for abnormality and for small vessel disease presence. Of these reports, 80% (<span style="font-style:italic;">n; = 1630) were used to fine-tune pretrained transformer models to classify scans. Five-fold cross-validation was used in model development. The models were tested on the remaining 20% of the reports (<span style="font-style:italic;">n; = 408). SHapley Additive exPlanations (SHAP) were used to explain the results.Results;The models exhibited excellent classification performance, with a mean receiver operating characteristic (ROC) area under the curve (AUC) of 0.98 for abnormality classification and 0.99 for small vessel disease classification. SHAP highlighted relevant words in both cases.Conclusions;This application validated the use of a pretrained transformer in detecting confounding data in research cohorts, and exhibited explainable results that allow the models’ decisions to be understood. By highlighting the specific report terms that drive each prediction, the explainable model output can be reviewed and critiqued by subject matter experts, supporting trust, error analysis, and iterative refinement of AI tools within clinical workflows.Advances in knowledge;This application demonstrates the feasibility of explainable report classification, and the fine-tuned model could be used in future for automatic removal of confounding data from radiology datasets, while providing transparent, case-level justifications that support audit, governance, and clinician acceptance.;

Objectives;Dynamic MRI of the upper vocal tract is increasingly used to study speech. Image segmentation is often required to analyse the organs of speech; however, manual segmentation is labour intensive and time consuming and automatic methods are being developed. In this paper, a new hybrid transformer network is proposed for such task.Methods;We introduce a deep learning-based decoder model termed “Progressively Refinement Decoding (PRORED).” This model incorporates a directional field (DF) module designed to capture the contour details of features. The acquired contour information is leveraged to refine the boundaries both between and within classes. By integrating the DF module at different stages of the decoder, features are enhanced progressively, ensuring a more detailed and accurate segmentation.Results;Our model is evaluated using a publicly accessible speech MRI dataset and a cardiac dataset. The metrics employed are the Dice coefficient and the Hausdorff distance. Results indicate that our model attains an average Dice coefficient of 97.78% and a Hausdorff distance of 6.84 mm. Additionally, our network was able to identify closure patterns more efficiently than the baseline network and previously published work. In addition, the model was also evaluated on a cardiac dataset, and achieved 91.90% dice score.Conclusions;The proposed model leads to a more accurate segmentation of speech MRI data and in particular allows for a better velopharyngeal closure study. The proposed model was also evaluated on a cardiac dataset and achieved competitive performance, showing its strong generalizability.Advances in knowledge;First model that utilizes vision transformer and progressive refinement decoder to segment dynamic speech MRI.;

<span class="paragraphSection">Abstract;Liver diseases consistently plague people’s daily lives as a result of their high morbidity and mortality rates. Ultrasound (US), favoured by its flexibility, free of radiation, cost-effectiveness, and real-time capabilities, has been commonly employed as one of the first-line imaging tools for hepatic conditions. Artificial intelligence (AI) algorithms are increasingly applied to automatically identify intricate patterns and perform quantitative analyses in US imaging, potentially reducing radiologists’ workload and improving diagnostic efficiency. AI-based US has been of substantial assistance in detecting, diagnosing, screening as well as monitoring of various liver diseases, and has attracted extensive attention among the medical community. In this review, we present a general introduction to AI in medical imaging; we next review its rapidly evolving applications in liver US, covering evaluation of hepatic steatosis severity, assessment of liver fibrosis, identification of focal hepatic lesions, preoperative prediction of high-risk pathological characteristics, assessment of postoperative prognosis, and the analysis of the model of integrated application of multi-omics data; finally, we present an outlook on the clinical applications of AI-based US in the liver diseases.;