Despite rapid advancements in artificial intelligence (AI) for medical imaging, widespread clinical adoption remains limited. In March 2025, the Academy for Radiology & Biomedical Imaging Research convened a cross-sector roundtable to examine operational and structural challenges in AI development and implementation. Researchers, department leaders, government representatives, and industry executives participated in a structured two-stage discussion using the AI lifecycle and a simplified failure modes and effects analysis (sFMEA) framework. In the first stage, attendees examined each phase of the AI lifecycle to identify domains where implementation barriers arise. In the second stage, mixed stakeholder groups applied a qualitative sFMEA approach to analyze process vulnerabilities within those domains and discuss mitigation approaches. This manuscript summarizes the session design, synthesizes key domains, and presents illustrative mitigation approaches across five areas: governance, use cases, implementation, cost, and regulation. The discussion identified recurring challenges related to fragmented priorities, infrastructure constraints, and regulatory complexity, as well as the need for clearer governance structures and more consistent evaluation processes to improve coordination across stakeholders.

Purpose To evaluate the feasibility of a transformer structure for fusing longitudinal multimodal radiomic features from chest radiographs (CXRs) and CT images to predict outcomes and identify associated clinical events in patients with COVID-19. Materials and Methods This retrospective study analyzed de-identified longitudinal CXRs and CT images in patients with polymerase chain reaction-confirmed COVID-19. Proprietary patient data (site 1) were collected between July 2020 and May 2021, and open-access patient data (obtained before February 1, 2020) were collected from site 2. Clinical outcomes included mortality, intensive care unit admission, and ventilator use during any follow-up visit. Radiomic features were extracted from lung regions in CXRs and CT images using the Cancer Imaging Phenomics Toolkit and integrated using a transformer-based model. Patient data were partitioned into training, validation, and test sets (ratio, 65:15:20). Subgroup analyses were performed across sex, site, and modality. Model performance was assessed using area under the receiver operating characteristic curve (AUC) and weighted AUC scores, with statistical significance assessed using Student t tests. Results The study included 2274 patients (946 from site 1, 1328 from site 2; mean age, 59.84 years ± 16.84, 1171 male patients). Weighted testing AUCs for predicting outcomes were 0.86 (95% CI: 0.85, 0.86) for mortality, 0.82 (95% CI: 0.81, 0.82) for intensive care unit admission, and 0.86 (95% CI: 0.86, 0.87) for ventilator usage, outperforming models trained solely on cross-sectional data or single-modal data (P < .05). Conclusion Transformer-based fusion of longitudinal multimodal radiomic data effectively predicted clinical outcomes and events associated with COVID-19. Keywords: COVID-19, Lung, Radiomics, Multi-Head Attention, Multimodal, Longitudinal, CT, Chest Radiography Supplemental material is available for this article. © RSNA, 2026.

Purpose To develop an explainable radiopathomic graph deep learning (RPGDL) system for multiscale spatial-contextual modeling of intratumoral heterogeneity and evaluate its performance for the prediction of pathologic complete response (pCR) to neoadjuvant therapy in breast cancer. Materials and Methods The RPGDL system was developed from dual-center retrospective analysis of patients with biopsy-proven invasive breast cancer (May 2018-August 2024). For each tumor, individual radiomic and pathomic graphs were generated from pretherapeutic MRI and hematoxylin-eosin-stained biopsy slide images, respectively. These graphs were then processed by three distinct graph neural networks (GNNs): radiomic, pathomic, and radiopathomic. GNN performance was assessed with the area under the receiver operating characteristic curve (AUC), net reclassification index (NRI), and integrated discrimination improvement (IDI). A multifaceted approach was used to explain the GNNs' predictions. Results The training set included 582 (mean age, 52 years ± 9 [SD]) patients and the external test set 468 (50 years ± 10) patients from centers 1 and 2, respectively. The radiomic GNN achieved AUCs of 0.89 (95% CI: 0.85, 0.93) in the training set and 0.84 (95% CI: 0.80, 0.89) in the external test set; the pathomic GNN achieved AUCs of 0.87 (95% CI: 0.83, 0.91) in the training set and 0.83 (95% CI: 0.78, 0.88) in the external test set, with no significant difference between them (P > .05). The radiopathomic GNN outperformed both single-modality GNNs (training set: AUC, 0.95 [95% CI: 0.92, 0.98]; external test set: AUC, 0.91 [95% CI: 0.87, 0.94]; P < .05; NRI and IDI confirmed). Pathomic graphs dominated probability increases for pCR predictions, while radiomic graphs drove probability decreases for non-pCR predictions. Multifaceted analyses verified GNNs' explainability. Conclusion The developed RPGDL system enabled multiscale spatial-contextual intratumoral heterogeneity modeling for high-performance, explainable prediction of pCR to neoadjuvant therapy in breast cancer. Keywords: Dynamic Contrast-enhanced MRI, Breast, Tumor Response, Radiology-Pathology Integration, Prognosis, Principal Component Analysis, Perception, Supervised Learning, Reconstruction Algorithm Supplemental material is available for this article. © RSNA, 2026.

Large language models (LLMs) are increasingly explored for radiology-related applications, yet their vulnerability to cognitive biases remains undercharacterized. The aim of this study was to investigate whether targeted prompts exploiting cognitive biases degrade LLM accuracy on radiology board-style questions. Ten contemporary LLMs were evaluated on 200 text-based and 200 multimodal American Board of Radiology examination-style questions under baseline and three cognitive bias prompts: authority bias prompts (ABPs), complexity bias prompts (CBPs), and anchoring bias prompts (AnBPs). Two mitigation approaches-a prompt bias audit and a one-shot mitigation strategy-were also evaluated. Under baseline prompts, models achieved a mean accuracy ± SD of 84.8% ± 5.5 (154-186 of 200) for text-based and 59.5% ± 7.7 (101-143 of 200) for multimodal questions. All models showed reduced accuracy to cognitively biased prompts, with ABP, CBP, and AnBP yielding absolute declines of 21.1%, 10.1%, and 4.4%, respectively, for text questions (P < .001 for each), and 44.9%, 44.4%, and 39.6%, respectively, for multimodal questions (P < .001 for each). The prompt bias audit increased accuracy by 5.6% for text-based and 15.8% for multimodal questions, whereas the one-shot mitigation yielded gains of 4.0% for text questions and 24.9% for multimodal questions. These findings demonstrate that LLMs are susceptible to cognitively biased inputs. Keywords: Technology Assessment, Use of AI in Education, Social Implications Supplemental material is available for this article. © RSNA, 2026 See also commentary by Tayebi Arasteh and Truhn in this issue.

Chest radiography is the most frequently performed imaging examination worldwide, and increasing demand has contributed to reporting delays in many health systems. This prospective multicenter silent trial evaluated the diagnostic performance of a commercially available artificial intelligence (AI) model for triaging normal chest radiographs across five National Health Service hospital sites in the United Kingdom over a 12-month period. A total of 63 083 adult chest radiographs were analyzed. The AI model classified 50 661 examinations (80%) as abnormal and 12 422 (20%) as normal. The model achieved 97% sensitivity, 35% specificity, 57% positive predictive value, and 94% negative predictive value for detecting abnormal chest radiographs. Expert review of discrepant cases, after exclusion of 412 natural language processing labeling errors, identified 31 clinically significant AI misses, corresponding to an estimated clinically significant miss rate of 0.05%. Most missed findings involved subtle or overlapping lesions. Concordance between the AI model and radiologist reports for normal examinations occurred in 18.5% of chest radiographs, indicating that nearly one-fifth of examinations could potentially be deprioritized for reporting. These findings suggest that AI-assisted triage of chest radiographs may help prioritize reporting workflows while maintaining a low rate of clinically significant missed findings, although further research is warranted to evaluate clinical implementation. Keywords: Artificial Intelligence, Chest Radiography, Computer-aided Diagnosis, Thorax Supplemental material is available for this article. © RSNA, 2026.

The introduction of foundational models, specifically large language models, has promised a health care transformation. However, the field is rapidly evolving toward autonomous agent systems, defined as artificial intelligence (AI) entities that perceive and react to their environment to achieve specific goals-representing a paradigm shift from passive information retrieval to proactive, goal-oriented clinical assistance. Agentic AI systems transcend static knowledge limitations through core capabilities including persistent memory systems that maintain context across patient encounters, knowledge retrieval tools connecting to medical repositories through retrieval-augmented generation techniques, and computer use functionality enabling navigation of clinical software interfaces. Agentic workflows introduce sophisticated coordination mechanisms including hierarchical, collaborative, and sequential patterns demonstrating superior performance compared with single-agent approaches. Multiagent systems can autonomously coordinate entire clinical workflows across the entire radiology life cycle, from preacquisition protocol optimization through initial image analysis, specialized tool deployment, and preliminary report generation. However, successful clinical deployment requires systematic consideration of complexity thresholds, economic sustainability, cybersecurity frameworks, bias mitigation strategies, and appropriate governance structures. Critical challenges include managing the probabilistic nature of underlying models within deterministic clinical workflows, ensuring adequate human supervision, and preventing overcomplication of established processes. A structured four-phase implementation roadmap addresses these considerations through incremental progression from low-risk automation to comprehensive workflow orchestration while maintaining rigorous safety standards. As foundation models advance and interoperability standards mature, agentic AI will reshape radiology practice paradigms. Success depends on resolving stakeholder responsibility questions while orchestrating technological capabilities with clinical accountability, ensuring autonomous systems augment rather than replace professional judgment in pursuit of improved patient outcomes. Keywords: Informatics, Named Entity Recognition, Patient Scheduling/No-Show Prediction, Resource Allocation, Impact of AI on Education, Artificial Intelligence, Large Language Models, Agentic AI, Multi-Agent Systems, Radiology Workflow, Clinical Decision Support, Health Care Automation © RSNA, 2026.

Purpose To develop a digitized integrated feature-based interpretable machine learning classification model to accurately recognize complex thyroid nodules while efficiently diagnosing conventional thyroid nodules (thyroid nodules with typical benign or malignant US features). Materials and Methods Thyroid US images depicting pathologically confirmed nodules were retrospectively collected from seven medical centers in China (January 2011-December 2021). An interpretable classification model consisting of two independent masks was developed and defined as "UltraMC." The front-end network was trained to identify conventional thyroid nodules using four digitized features, and the back-end network collected nodules classified as benign in the previous framework for secondary analysis to clarify their final diagnosis. UltraMC performance was evaluated using accuracy, sensitivity, specificity, and confusion matrices. Results The total dataset included 73 826 patients with thyroid US images (mean age, 45.56 years ± 11.21 [SD]; 54 398 female). Diagnostic accuracy of the front-end network for detecting conventional thyroid nodules was 92.9% (13 718 of 14 765), and accuracy of the back-end network for classifying mummified thyroid nodules (MTNs) was 88.5% (652 of 737). The overall diagnostic accuracy of the US MTN classification model (UltraMC) was 91.8% (14 228 of 15 502). The areas under the receiver operating characteristic curve of the front-end network and UltraMC in identifying conventional thyroid nodules were 0.98 (95% CI: 0.98, 0.98) and 0.96 (95% CI: 0.96, 0.97), respectively. Conclusion The proposed two-layer interpretable classification model achieved high diagnostic accuracy for both conventional and mummified thyroid nodules. These findings demonstrate that digitized US features integrated into a white box framework can effectively support classification of complex thyroid nodules. Keywords: Ultrasound, Head/Neck, Thyroid, Diagnosis, Convolutional Neural Network (CNN), K-Means, Random Forest, Thyroid Nodule, Interpretable, Digital, Mummified Thyroid Nodules Supplemental material is available for this article. © RSNA, 2026.

Purpose To develop and validate a deep learning (DL) model for automated assessment of coronary stenosis in vessels with heavily calcified plaques at coronary CT angiography (CCTA), using quantitative coronary angiography as the reference standard. Materials and Methods A total of 10 101 CCTA examinations (June 2017-December 2020) from three tertiary hospitals in China were retrospectively collected for DL model development. External testing dataset 1 included 442 CCTA examinations (Agatston score > 300) from two independent hospitals (January 2021-May 2022) for performance evaluation. The separate external testing dataset 2 of 120 CCTA examinations was used for a reader study assessing whether DL assistance improved diagnostic accuracy among junior, attending, and senior radiologists. External testing dataset 3 included 150 prospectively collected CCTA examinations (June-July 2023) that were analyzed to compare model performance against clinical reports, simulating real-world deployment. Model diagnostic performance was assessed using receiver operating characteristic analysis, with quantitative coronary angiography as the reference. Results In external testing dataset 1, specificities for detecting 50% or more stenosis were 78%, 72%, and 48% and the areas under the receiver operating characteristic curve (AUC) were 0.89, 0.90, and 0.87 at the segment, vessel, and patient levels, respectively. In external testing dataset 2, DL assistance improved radiologist specificity by 7%-11% (P < .001) with improving AUC and increased interreader agreement (Δκ = 0.155-0.228; P < .05). In external testing dataset 3, the model demonstrated 53% specificity and a higher AUC versus clinical reports (0.91 vs 0.76; P < .001). Conclusion The proposed DL model accurately detected coronary stenosis of heavily calcified plaques at CCTA and improved diagnostic performance of radiologists. Keywords: CT Angiography, Cardiac, Heart, Arteriosclerosis, Calcifications, Calculi, Quantification, Diagnosis Supplemental material is available for this article. © The Author(s) 2025. Published by the Radiological Society of North America under a CC BY 4.0 license. See also commentary by Maiter and Alabed in this issue.

Purpose To develop a self-supervised text-vision framework to detect abnormalities on brain MRI scans by leveraging free-text neuroradiology reports, eliminating the need for expert-labeled training datasets. Materials and Methods This retrospective and prospective multicenter study included 81 936 brain MRI examinations and corresponding radiology reports for adult patients at two UK National Health Service hospitals from January 2008 to December 2019 for training and internal testing and 1369 prospectively collected examinations between March 2022 and March 2024 from four separate National Health Service hospitals for external testing (ClinicalTrials.gov no. NCT04368481). A neuroradiology language model (NeuroBERT) was trained using self-supervised tasks to generate report embeddings. Convolutional neural networks (one per MRI sequence) were trained to map scans to embeddings by minimizing mean squared error loss. The framework then detected abnormalities in new examinations by scoring scans against query sentences using text-image similarity. Model diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUC). Results The framework achieved an AUC of 0.95 (95% CI: 0.94, 0.97) for normal versus abnormal classification and generalized to external sites with examination-level AUCs of 0.90 (95% CI: 0.86, 0.93) in Bedford, 0.87 (95% CI: 0.83, 0.90) in Nottingham, 0.86 (95% CI: 0.83, 0.90) in Norwich, and 0.85 (95% CI: 0.81, 0.89) in Yeovil. In five zero-shot classification tasks-acute stroke, multiple sclerosis, intracranial hemorrhage, meningioma, and hydrocephalus-the framework achieved a mean AUC of 0.89 (range, 0.77-0.93). For visual-semantic image retrieval, mean precision was 0.84 among the top 15 images across seven pathologies. Conclusion The self-supervised text-vision framework accurately detected brain MRI abnormalities without expert-labeled datasets. Clinical trial registration no. NCT04368481 Keywords: Head and Neck, Unsupervised Learning, Convolutional Neural Network (CNN), Neuroradiology © The Author(s) 2025. Published by the Radiological Society of North America under a CC BY 4.0 license. Supplemental material is available for this article. See also commentary by Ghodasara in this issue.