Erickson, Bradley J.

ABSTRACT Background Family history of pancreatic cancer (PC), pathogenic germline variants (PGVs), and increased intrapancreatic fat (IPF) are individually associated with increased PC risk. Whether IPF is elevated in high‐risk individuals (HRIs) with PC family history or PGVs is unknown. Although MRI is the preferred modality for IPF measurement, region‐of‐interest–based approaches are prone to sampling bias because pancreatic fat is heterogeneously distributed. We aimed to compare IPF among HRIs and matched non‐HRI controls using an Artificial intelligence (AI)‐based tool for whole pancreatic IPF measurement applied to magnetic resonance imaging (MRI). Methods We developed and validated a deep‐learning pipeline for IPF measurement that incorporates automated pancreas segmentation on T1‐weighted MRI images registered to the corresponding iterative decomposition of water and fat with echo asymmetry and least squares estimation images to compute IPF using proton density fat fraction (PDFF). This pipeline was used to quantify IPF and assess IPF distribution patterns in 85 HRIs and 170 non‐HRIs matched 1:2 for age, sex, BMI, and diabetes status. A surface erosion approach was developed to minimize the impact of peripancreatic fat on IPF measurement. Linear models adjusting for matching variables were used to assess associations. Results Model‐expert agreement for whole‐pancreas PDFF was excellent (intraclass correlation coefficient 0.99; 95% CI, 0.98–0.99 mean absolute error 1.4%). Mean PDFF did not differ between HRIs and controls (25.8% [11.5] vs. 26.9% [11.4]; p = 0.32). PDFF increased with age ( p = 0.015), male sex ( p < 0.001), presence of diabetes ( p = 0.009), BMI ( p < 0.001). IPF distribution patterns did not differ by risk group. Conclusion An AI‐based approach provides accurate, scalable, and objective assessment of IPF quantity and distribution. IPF quantity and spatial distribution do not appear to differ based on germline or familial predisposition to PC.

Artificial intelligence can be applied across radiology education to enhance curriculum development and delivery through data-driven needs assessment, content and assessment generation, personalized learning, and program evaluation and management.

Abstract Background & Objectives Cholangiocarcinoma (CCA) is a rare and aggressive malignancy of the bile ducts characterized by complex imaging patterns, high inter-reader variability, and limited subspecialty expertise. Despite the critical importance of early detection for improving patient outcomes, most artificial intelligence (AI) tools developed for radiology have focused on other organ systems, leaving CCA largely underserved. Furthermore, clinicians with relevant domain expertise frequently lack the programming background required to participate meaningfully in AI development. This work aimed to create an accessible, educational AI framework “AI Starter” designed to empower radiologists and clinicians to initiate medical image analysis projects for CCA without requiring extensive coding knowledge. Methods A structured, end-to-end AI pipeline was introduced covering the key stages of image acquisition, preprocessing, segmentation, feature extraction, and predictive modeling. Open-source frameworks, specifically MONAI and PyTorch, were employed to ensure reproducibility and accessibility. Transfer learning strategies were incorporated to address the challenge of limited CCA-specific training datasets. The pipeline was designed with modular components to reduce technical barriers and accommodate users across varying levels of computational experience. Results Radiologists and clinicians were able to successfully engage with and deploy AI tools for CCA imaging analysis using this framework. Transfer learning demonstrably improved model performance under data-limited conditions, and the open-source implementation supported reproducible research workflows. Key challenges identified included scanner variability, segmentation instability, feature instability, and model overfitting, all of which were transparently mapped within the pipeline. Conclusion The AI Starter framework offers a practical and educational entry point for the CCA community including basic scientists and clinicians to meaningfully engage with and understand AI-driven imaging analysis.

Abstract Glioblastoma (GBM), isocitrate dehydrogenase wild-type (IDHwt) and central nervous system diffuse large B-cell lymphoma (CNS-DLBCL) are aggressive brain tumors with overlapping MRI features, yet distinct treatment approaches. Noninvasive tools are needed to aid in differential diagnosis. Deep learning on T1 postcontrast and T2-weighted MRI sequences were used to differentiate GBM and CNS-DLBCL. A three-stage temporal study design was utilized. Model development was performed on 146 patients with CNS-DLBCL and 146 age-matched, sex-matched, and MRI year–matched patients with GBM diagnosed at Mayo Clinic between 1998 and 2019. Models were tested on independent temporal test cohorts. Initial testing included 240 independent GBM diagnosed at Mayo Clinic between 1998 and 2019. The prospective test cohort included 37 patients with CNS-DLBCL and 256 patients with GBM diagnosed at Mayo Clinic after January 1, 2020, and 36 patients with CNS-DLBCL diagnosed at an external institution. Of the patients diagnosed at Mayo Clinic, 47% had MRIs generated from non-Mayo institutions. Two different model approaches were compared: (i) ensemble approach using area under the receiver operating characteristic curve (AUC) and cross-validation for model selection and (ii) loss approach minimizing cross-entropy loss and cross-validation to evaluate prediction performance. The AUCs on the prospective test cohort were 0.84 [95% confidence interval (CI), 0.78–0.90] and 0.83 (95% CI, 0.77–0.88) for the ensemble and loss approaches, respectively. Stability of ensemble prediction improved with the increasing number of models. Stratified AUC analysis demonstrated consistent performance across sex and age. We utilized a robust temporal study design and applied 2 different analytic approaches to develop a classification model. The findings confirm the feasibility of using MRI-based deep learning models to differentiate GBM from CNS-DLBCL. Significance: GBM, IDHwt and CNS-DLBCL are aggressive brain tumors with overlapping MRI features, yet distinct treatment approaches. Noninvasive tools are needed to aid in differential diagnosis. We developed MRI-based deep learning models to differentiate GBM, IDHwt from CNS-DLBCL using a rigorous three-stage temporal design that included prospective validation. The model AUC on a prospective cohort was 0.84.

The field of radiology is experiencing rapid adoption of large language models (LLMs), yet their tendency to generate hallucinations (plausible but incorrect information) remains a significant barrier to trust. This comprehensive review evaluates emerging agentic artificial intelligence (AI) approaches, including multi-agent role-based systems, retrieval-augmented generation (RAG), and uncertainty quantification, to assess their potential for reducing hallucinations in radiology workflows. Evidence from 2024 to 2025 demonstrates that agentic AI can improve diagnostic accuracy and reduce error rates, though these methods remain computationally demanding and lack comprehensive clinical validation. Multi-agent frameworks enable cross-validation through role-based specialization and systematic workflow orchestration, while RAG strategies enhance accuracy by grounding responses in verified medical literature. Within multi-agent systems, uncertainty quantification enables agents to communicate confidence levels to one another, allowing them to appropriately weigh each other’s contributions during collaborative analysis. While multi-agent frameworks and RAG strategies show significant promise, practical deployment will require careful integration with human oversight, robust evaluation metrics tailored to medical imaging tasks, and regulatory adaptation to ensure safe clinical use in diverse patient populations and imaging modalities.

Abstract Glioblastoma, IDH-wildtype (GBM), and primary central nervous system lymphoma (PCNSL) are two aggressive brain tumors with distinct treatment strategies but overlapping imaging features on MRI. Accurate preoperative differentiation remains a clinical challenge, due to their overlapping imaging features, even when advanced MRI techniques are employed. This study explored deep learning approaches leveraging multi-sequence MRI to distinguish GBM from PCNSL. Models were developed utilizing 146 PCNSL cases, which were matched to146 GBM by age, sex and year of diagnosis. A 3D-DenseNet121 convolutional neural network was trained using two analytical strategies: (i) loss-based model selection and (ii) ensemble prediction using five-fold cross-validation. Model performance was evaluated on an independent validation set of 256 GBM and 37 PCNSL from the same institution as well as 36 PCNSL cases from an external institution. All MRIs were preprocessed using the FeTS pipeline. Using the independent validation cohort, area under the receiver operating characteristic curve (AUC) and 95% confidence intervals (CI) were computed overall and stratified by demographic characteristics. Using both T1 post-contrast (T1Gd) and T2-weighted images, the loss-based model achieved AUC=0.83 (95% CI: 0.77-0.88) and the ensemble model yielded an AUC=0.82 (95% CI: 0.76-0.89). Model performance was similar across males and females; however, model performance was better for subjects older than 50 years of age versus younger than 50. When using only T1Gd images, the loss-based model dropped to AUC=0.55 (95% CI: 0.48-0.63), whereas the ensemble model had AUC=0.79 (95% CI: 0.72-0.85). Models trained with only T2 images achieved AUCs of 0.67 (95% CI: 0.60-0.73) (loss-based) and 0.75 (95% CI: 0.69-0.82) (ensemble). These results highlight the added value of the multi-sequence input. Our results confirm the feasibility of using automated deep learning models to differentiate GBM from PCNSL. Notably, the model generalized well to external MRI data, emphasizing its potential utility as a diagnostic aid in real-world clinical workflows.

Introduction: Carotid artery intraplaque hemorrhage (IPH) is increasingly recognized as a predictive marker of cerebrovascular events. Despite widespread utilization of contrast-enhanced CT angiography (CTA) for the evaluation of carotid atherosclerosis, diagnosis of IPH on CTA is challenging, limiting risk stratification capabilities. We evaluated convolutional neural networks (CNN) for detection of histology-proven IPH in carotid CTA. Methods: The study included carotid endarterectomy patients from our center who had preoperative CTA and available plaque specimens from the Mayo Clinic Carotid Artery Atherosclerosis Trial (MCAT) Biobank. A validated segmentation model localized common and internal carotid arteries (CCA/ICA). CCA bifurcation served as landmark after a large language model, Llama3.0-70B, verified culprit plaque proximity to the bifurcation from CTA reports. A 3×3×4cm box around ipsilateral CCA bifurcation was sampled from CTA volumes and images were labeled IPH by histology study of carotid specimen (Fig 1). Given limited cases with concurrent imaging and histology, we employed a robust cross-validation strategy: 22 cases (20%) held-out for testing, with remaining data 86 cases used for 5-fold cross-validation (stratified by IPH presence). The five resulting models independently generated IPH probability scores, and ensembled predictions calculated by three methods to enhance generalizability: Majority and weighted voting based on binary IPH class and soft voting based on IPH probability scores. Results: Of 418 patients with histological analysis, 130 had CTA studies available. After excluding 22 cases lacking histological consensus, 108 cases were analyzed. In the 22 test cases the mean age was 68.1±8.5 years and 59% were male; In the cross-validation folds age ranged from 67.1±6.9 to 75.3±7.6 years 56-82% were male. Individual models' performance on the test set showed mean ROC-AUC of 0.79 (95% CI: ±0.05) and average precision of 0.85 (±0.05) (Fig 2). The ensemble approach achieved superior performance with soft voting demonstrated 82% ROC-AUC, 77% accuracy, 79% precision, 85% sensitivity, and 82% F1-score for IPH detection, outperforming majority voting and weighted voting (Fig 3). Conclusion: Automated IPH detection in widely available CTA imaging can turn it to a powerful stroke risk assessment tool, supporting personalized prevention strategies and surgical decision-making, ultimately reducing cerebrovascular morbidity and mortality.

Radiology is undergoing a paradigm shift from traditional single-function AI systems to sophisticated multi-agent networks capable of autonomous reasoning, coordinated decision-making, and adaptive workflow management. These agentic AI systems move beyond simple pattern recognition to encompass complex radiological workflows including image analysis, report generation, clinical communication, and care coordination. While multi-agent radiological AI promises enhanced diagnostic accuracy, improved workflow efficiency, and reduced physician burden, it simultaneously amplifies the long-standing “black box” problem. Traditional explainable AI methods, which are adequate for understanding isolated diagnostic predictions, fail when applied to multi-step reasoning processes involving multiple specialized agents coordinating across imaging interpretation, clinical correlation, and treatment planning. This paper examines how agentic AI systems in radiology create “compound opacity” layers of inscrutability from agent interactions and distributed decision-making processes. We analyze the autonomy–transparency paradox specific to radiological practice, where increasing AI capability directly conflicts with interpretability requirements essential for clinical trust and regulatory oversight. Through examination of emerging multi-agent radiological workflows, we propose frameworks for responsible implementation that preserve both diagnostic innovation and the fundamental principles of medical transparency and accountability.