RPS 1417 - Hot Topic: AI in emergency imaging

Author Block: T. Clement¹, L. Mabit¹, G. Davy¹, G. D'Assignies², R. Guillevin¹, G. Herpe¹; ¹Poitiers/FR, ²Nantes/FR
Purpose: AI-driven triage tools are increasingly used in emergency radiology to help detect and prioritize urgent conditions such as brain injuries, pneumothorax, and incidental pulmonary embolism. This study assesses the real-world impact of deploying multiple AI alerting systems.
Methods or Background: This retrospective multi-centric study was conducted over two months in three ER department. AI algorithms were applied to detect intracranial hemorrhage on CT, pneumothorax on chest X-ray, and incidental pulmonary embolism on contrast-enhanced CT across 2,336 CT and 119 Chest X-rays. Discrepancies between AI outputs and radiology reports were first reviewed by a radiologist and then resolved by an emergency radiologist. Performance was evaluated using diagnostic metrics.
Results or Findings: ICH detection : Among 682 head CT scans, AI flagged 133 positives(19.5%,133/682), including 46 false positives (6.7%,46/682) and 11 false negatives(1.6%,11/682). AI detected 2 additional true positives(1.5%,2/133) missed by radiologists.
Pneumothorax Detection: Out of 119 chest X-rays, AI identified 3 positive cases(2.5%,3/119), including 1 missed by the radiologist (0.8%,1/119), with no false positives.
Incidental Pulmonary Embolism Detection : In 1654 contrast-enhanced CT scans, AI flagged 70 positives(4.2%,70/1654), with 22 true positives(31.4%,22/70) and 9 missed by radiologists(41%,9/22). AI increased the detected prevalence from 0.8%(14/1654) to 1.3%(22/1654), but the false-positive rate was 68.6%(48/70).
Overall, at the cost of about 1.5 false alerts per day, AI helped uncover one life-threatening condition every five days.
Conclusion: AI triage tools in emergency radiology enhance detection of missed critical findings and remain valuable if false positives are carefully managed to avoid alert fatigue.
Limitations: This retrospective, short-duration study limits generalizability, especially given the small sample size for some modalities. The reference standard relied solely on radiologist re-reads without follow-up confirmation. Finally, AI real impact on workflow, alert fatigue, and patient outcomes was not assessed.
Funding for this study: No funding for this study
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: J. WU¹, Y. Li², P. Zhang¹, H. Huang¹, J. Hu¹; ¹Fuzhou/CN, ²Zhangzhou/CN
Purpose: To identify clinical predictors and develop a prediction model to identify non-traumatic pediatric patients at risk of abnormal head CT findings in emergency department(ED).
Methods or Background: Emergency pediatric patients (≤14 years old) were identified from four tertiary general hospitals. Age; Gender; Medical history; Fever;Crying; Seizure; Headache; Dizziness; Syncope; Vomiting; Abnormal Physical examination; Impaired Consciousness, etc., were used as candidate clinical factors. Multivariate logistic regression analysis were used to identify the independent clinical predictors of abnormal head CT findings. Datasets from two hospitals were used for model training and the other two for external validation. We developed models using logistic regression and machine learning (KNN, NNet, Random Forest, XGBoost, Naive Bayes).The best-performing model was identified, assessed with Decision Curve Analysis (DCA), and its feature contributions were explained via SHAP values.
Results or Findings: 127 of 2,272 ( 5.6%) cases with abnormal head CT findings. Younger age (OR=0.89), medical history (OR=40.43), non fever(OR=0.35), crying(OR=11.13), vomiting (OR=4.75), headache(OR=2.18), physical examination abnormalities (OR=23.17), and consciousness disorders (OR=21.69) were clinical independent predictors. XGBoost was the top-performing model, with validation AUCs of 0.84 (Test1) and 0.88 (Test2). It showed a balanced sensitivity and specificity(Test1: 67.9% sensitivity, 91.0% specificity; Test2: 81.2% sensitivity, 81.6% specificity) at the Youden threshold and provided the highest net benefit on DCA in the clinically relevant range (0.20-0.30), reducing interventions by 57.8–90.6 per 100 patients. SHAP analysis showed impaired consciousness, age, fever, abnormal physical examination, and seizure were the top predictors of abnormal head CT findings.
Conclusion: The validated XGBoost model enables more selective uses of head CT for emergency pediatric patients, potentially reducing unnecessary scans without a loss in sensitivity.
Limitations: The relatively small number of abnormal CTs may affect the stability of the estimates for some predictor variables.
Funding for this study: Joint Funds for the innovation of science and Technology, Fujian  province(Grant number : 2024Y96020147)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical approval of this study was obtained from the Ethics Committee of these four hospitals（Fujian Provincial Hospital, The First Affiliated Hospital of Fujian Medical University, Fujian Medical University Affiliated Zhangzhou Hospital and the 900th Hospital of the PLA Joint Logistic Support Force ）

Author Block: E. E. AKTUFAN; Antalya/TR
Purpose: To evaluate the real-world performance of an AI-powered intracranial hemorrhage (ICH) detection system in emergency head CT. After the session, participants will understand the diagnostic strengths and limitations of AI, recognize common causes of false results, and appreciate its implications for clinical workflows.
Methods or Background: Deep learning algorithms show high accuracy in ICH detection, but real-world validations are limited. This retrospective study included all emergency non-contrast head CTs performed over three months at a tertiary hospital. Cases with motion and metallic artifacts were retained to reflect real practice. The ground truth was defined by board-certified radiologist reports, which were retrospectively compared with AI outputs (hStroke V1, Hevi AI, Istanbul). The model employs a CNN-RNN architecture with attention for ICH subtype classification.
Results or Findings: In 1,421 CTs from 1,379 patients (mean age: 50.6 years; 57% male), 73 were hemorrhage-positive (5.8%). The AI achieved 97.5% accuracy, 95.9% sensitivity, 97.6% specificity, and 68.6% precision. Common subtypes were IPH (n=35), SAH (n=37), and SDH (n=39). Multiple subtypes occurred in 37 cases, most often IPH+SAH+SDH. Three hemorrhages were missed, while 24 false positives mainly arose from artifacts, masses, and calcifications.
Conclusion: The AI tool demonstrated high accuracy and sensitivity for ICH detection in a real-world emergency setting. False positives remain a concern, emphasizing radiologist oversight and improved artifact handling.
Limitations: This single-center retrospective study relied on board-certified radiologist reports as ground truth, without interobserver analysis. Subtype-specific stratification was limited by sample size. Only axial CTs were analyzed, excluding multiplanar reconstructions that may improve detection in anatomically complex regions.
Funding for this study: This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study protocol was approved by the institutional ethics committee (Protocol No: 2024-122, Decision No: 11/28; Approval Date: August 8, 2024). Informed consent requirements were waived due to the retrospective nature of the study and the anonymization of all datasets

Author Block: P. Reschke, K. Eichler, T. Vogl, C. Booz; Frankfurt/DE
Purpose: Accurate detection of intracranial hemorrhage (ICH) on non-contrast CT is critical in emergency settings, where missed diagnoses may delay treatment and worsen outcomes. While artificial intelligence (AI) models demonstrate high standalone performance, their additive value as a second reader for radiology residents is not well established.
Methods or Background: This retrospective study included 1,337 non-contrast head CT scans from 2015–2019 (670 ICH-positive, 667 ICH-negative). A previously validated AI model was used for ICH detection. Two radiology residents reviewed all scans in consensus, first without and later with AI support after a 30-day washout. Ground truth was established by expert consensus. Diagnostic performance metrics were calculated.
Results or Findings: AI assistance significantly improved radiology residents’ diagnostic performance. Sensitivity increased from 0.85 to 0.94 and specificity from 0.87 to 0.97 (both p < 0.01). ROC-AUC rose from 0.86 to 0.95, and PR-AUC from 0.83 to 0.95 (p < 0.0001). The number of false negatives dropped from 101 to 41 with AI support. The greatest benefit was observed in subdural hematomas (SDH), where misses declined from 32 to 9 (20.3% to 5.7%; p < 0.001), corresponding to a 72% relative risk reduction. Misses also decreased for intraparenchymal hemorrhages (IPH: 37 to 20; RRR 46%) and subarachnoid hemorrhages (SAH: 30 to 11; RRR 63%). AI support reduced common error sources: small hemorrhage volume (48 to 21), atypical locations (30 to 12), and image-degrading artifacts (23 to 8). False positives fell from 87 to 21.
Conclusion: By reducing diagnostic errors and supporting learning, AI serves as a valuable second reader for radiology residents—enhancing both patient safety and resident training in ICH detection.
Limitations: This study did not evaluate radiologists’ trust in or interaction with the system.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical approval was granted by the institutional review board of the University of Frankfurt and written informed consent was waived due to the retrospective nature of the study (approval number: 19-236).

Author Block: D. Ferreira Branco, P. Botti, A. Platon, P-A. A. Poletti, S. Boudabbous; Geneva/CH
Purpose: To evaluate a commercial AI fracture detection tool on foot/ankle radiographs, with emphasis on midfoot fractures (Chopart/Lisfranc), compared with board-certified musculoskeletal radiologists using a composite reference standard (CBCT and/or clinical follow-up).
Methods or Background: This retrospective single-center study included all emergency radiographs for adult patients with foot/ankle trauma over six months. Radiographs were first interpreted in routine workflow by radiologists, then independently (stand alone) by the AI tool (BoneView™, Gleamer), each blinded to the other. Descriptive analyses summarized fracture prevalence and distribution (overall, Chopart, Lisfranc). Diagnostic performance of AI and radiologists was assessed, with inter-reader agreement (Cohen’s κ). Paired comparisons used McNemar’s test.
Results or Findings: In total, 701 studies were analyzed; 319 fractures (45.5%) were found: 24 Chopart (7.6%), 22 Lisfranc (6.8%), and 273 (85.6%) other bony structures. AI achieved overall sensitivity 74.3%, specificity 83.0%, accuracy 79.0%. Radiologists achieved sensitivity 84.0%, specificity 95.5%, accuracy 90.3% (p=0.145); κ=0.65.
For Chopart fractures, AI sensitivity was 62.1%, specificity 99.6%, accuracy 98.0%; radiologists achieved sensitivity 82.8%, specificity 99.7%, accuracy 99.0% (p=0.180); κ=0.80. For Lisfranc fractures, AI sensitivity was 65.4%, specificity 99.9%, accuracy 98.6%; radiologists achieved sensitivity 80.8%, specificity 100.0%, accuracy 99.3% (p=0.453); κ=0.82
Conclusion: AI and radiologists achieved comparable overall performance for detecting foot/ankle fractures on radiographs, with substantial agreement. Radiologists showed a better performance in detecting Lisfranc/Chopart fractures than AI. Our findings support the AI use as complementary help for radiologists in a busy workflow, with targeted attention to midfoot injuries where detection remains challenging.
Limitations: Given the low prevalence of Chopart and Lisfranc fractures, the study period was relatively short.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical Board approval (CER 2020-02812)

Author Block: H-L. Kaatsch¹, D. Dillinger¹, C. Bauer¹, D. Friedmann², B. Schmidt², S. Waldeck¹, D. Overhoff¹; ¹Koblenz/DE, ²Forchheim/DE
Purpose: This study aims to assess a prototype hand unfolding reformation algorithm in polytrauma whole-body photon-counting CT (WB-PCCT) regarding time saving and image quality compared to a conventional bone reconstruction algorithm.
Methods or Background: We retrospectively analyzed 20 polytraumatized patients with bone injuries of the hand, who underwent polytrauma WB-PCCT, and reconstructed a total of 68 hands (matching 34 unfolded and 34 conventional (Br60 kernel) reconstructions). The hand unfolding reformation algorithm uses AI to automatically locate the hands in the WB-PCCT. After a deep learning-based detection of the course of the finger bones, the algorithm maps the original image to a predefined template using a deformation algorithm. This effectively unfolds each hand and enables visualization in the standard anatomical position. Quantitative image analysis was performed based on the calculation of the contrast-to-noise ratio (CNR). Two readers randomly assessed hand reconstructions with regard to reading time, subjective image quality (overall image quality, image noise perception, fracture delineation, fracture/joint dislocation) and diagnostic confidence using a 5-point Likert scale.
Results or Findings: Unfolded hand reformations significantly reduced (p<0.001) reading time for sufficient evaluation compared to conventional hand reconstructions by 57% (mean: 59 seconds versus 138 seconds). The novel reformation algorithm significantly improved CNR (28.18 versus 20.93; p <0.001). Subjective evaluation revealed comparably higher ratings for overall image quality, image noise perception, fracture and joint dislocation, and diagnostic confidence for unfolded hand reformations with equal ratings for fracture delineation.
Conclusion: The novel hand unfolding reformation algorithm for polytrauma WB-PCCT applied in this study led to 57% reduction in time required for hand assessment in time-critical polytrauma setting and improves both objective as well as subjective image quality compared to conventional bone reconstruction.
Limitations: Small sample size (n=20).
Retrospective design.
Funding for this study: No funding.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The present study was approved by the Local Ethics Committee of the Chamber of Physicians Rhineland-Palatinate in Mainz, Germany, and conducted in accordance with the Declaration of Helsinki.

Author Block: E. C. Stamoulaki, E. Detorakis, N. Christodoulides, D. Grigoropoulou, M. Klontzas, M. Raissaki; Heraklion/GR
Purpose: To assess performance of chest radiography without and with AI assistance compared to CT for abnormality detection in the emergency setting.
Methods or Background: 150 patients (males 63,33%), aged 17-92 years (mean 72,5) underwent frontal PA or AP and lateral radiographs and CT within 10 days (mean 1 day). A junior radiologist assessed radiographs for lung, mediastinal and pleural/rib abnormalities, initially without AI assistance (reading 1), and following a three-week weaning interval with AI assistance (reading 2). CT findings were subsequently recorded by another junior and a senior radiologist, formulating ground truth. Lesions were marked as 0=not present, 1=present. Statistical analysis was performed using R programming language (v4.2.2 in RStudio for MacOS). Agreement between CT and each reading was assessed with weighted kappa statistics. Confusion matrices were used to calculate sensitivity, specificity, balanced accuracy, positive and negative predictive value. P<0.05 was statistically significant.
Results or Findings: Agreement between radiography and CT for presence of abnormalities ranged from slight to moderate (weighted k-values 0.18-0.67). The addition of AI did not increase the sensitivity of radiographs for any of the assessed abnormalities. The weighted kappa ranged from slight for mediastinal masses (k=0.17 without vs 0.17 with AI) to substantial for pleural effusion (k=0.67 without vs 0.69 with AI) (P<0.001). Both readings showed excellent specificity (range 87.2%-100%). Both readings had similar sensitivity which was low for lung and mediastinal masses, nodules and rib fractures (13.3%-33.9%), moderate for consolidation (46.7% without AI vs 54.8% with AI) and high for pleural effusion (80% for both readings).
Conclusion: In our small cohort, the addition of AI did not result in a statistically significant improvement of agreement between radiographic findings and CT.
Limitations: Small sample size.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics committee of University General Hospital of Heraklion

Author Block: D. Grigoropoulou, E. C. Stamoulaki, E. Detorakis, N. Christodoulides, M. Klontzas, M. Raissaki; Herakleion/GR
Purpose: To investigate the educational effect of artificial intelligence on junior radiologists who report chest radiographs.
Methods or Background: Two sets of 75 patients who presented in the emergency department underwent PA or AP chest radiography and lateral chest radiography on the same day and chest CT within 0-10 days. The first set were assessed by one junior resident radiologist (reader) for lung, mediastinal and pleural/rib abnormalities separately without and with AI assistance. The second set was subsequently assessed by the same reader, blinded to CT-findings, which were recorded by two additional radiologists in consensus. Lesions were marked as 0=absent, 1=potentially/definitely present.
Statistical analysis was performed using R programming language (v4.2.2 in RStudio for MacOS). Agreement between CT and each reading was assessed with weighted kappa statistics. Confusion matrices were used to calculate sensitivity, specificity, balanced accuracy, positive and negative predictive value, which were compared between two sets. P<0.05 was statistically significant.
Results or Findings: Assessment of 1st set of radiographs without AI yielded a low to moderate agreement with CT, with a weighted kappa ranging between 0.108 (detection of mediastinal masses) and 0.529 (detection of pleural effusion). This agreement did not change with the assistance of AI, except for pleural effusion detection where agreement increased from 0.529 to 0.689. Assessment of the 2nd set without AI exhibited increased agreement with CT for nodules (0.131 before vs 0.478 after), lung (0.251 before vs 0.681 after) and mediastinal masses (0.108 before vs 0.520 after). Regarding sensitivity of radiographs for nodules, lung and mediastinal masses, it increased by 24-33%. Specificity was similar before and after the use of AI.
Conclusion: The addition of AI in chest radiography may increase sensitivity and may accelerate education of junior radiologists.
Limitations: Small sample size.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics committee of University General Hospital of Heraklion