Purpose: This study explores the utility of the large language models, GPT-3 and GPT-4, for in-depth patient education prior to interventional radiology procedures. Further, differences in answer accuracy between the models were assessed. Materials and methods: A total of 133 questions related to three specific interventional radiology procedures (Port implantation, PTA and TACE) covering general information as well as preparation details, risks and complications and post procedural aftercare were compiled. Responses of GPT-3 and GPT-4 were assessed for their accuracy by two board-certified radiologists using a 5-point Likert scale. The performance difference between GPT-3 and GPT-4 was analyzed. Results: Both GPT-3 and GPT-4 responded with (5) “completely correct” (4) “very good” answers for the majority of questions ((5) 30.8% + (4) 48.1% for GPT-3 and (5) 35.3% + (4) 47.4% for GPT-4). GPT-3 and GPT-4 provided (3) “acceptable” responses 15.8% and 15.0% of the time, respectively. GPT-3 provided (2) “mostly incorrect” responses in 5.3% of instances, while GPT-4 had a lower rate of such occurrences, at just 2.3%. No response was identified as potentially harmful. GPT-4 was found to give significantly more accurate responses than GPT-3 (p = 0.043). Conclusion: GPT-3 and GPT-4 emerge as relatively safe and accurate tools for patient education in interventional radiology. GPT-4 showed a slightly better performance. The feasibility and accuracy of these models suggest their promising role in revolutionizing patient care. Still, users need to be aware of possible limitations. Graphical

This study aimed to assess the feasibility of GPT-4 for answering questions related to contrast media with and without the context of the European Society of Urogenital Radiology (ESUR) guideline on contrast agents. The overarching goal was to determine whether contextual enrichment by providing guideline information improves answers of GPT-4 for clinical decision-making in radiology. A set of 64 questions, based on the ESUR guideline on contrast agents mirroring pertinent sections, was developed and posed to GPT-4 both directly and after providing the guideline using a plugin. Responses were graded by experienced radiologists for quality of information and accuracy in pinpointing information from the guideline as well as by radiology residents for utility, using Likert-scales. GPT-4's performance improved significantly with the guideline. Without the guideline, average quality rating was 3.98, which increased to 4.33 with the guideline (p = 0036). In terms of accuracy, 82.3% of answers matched the information from the guideline. Utility scores also reflected a significant improvement with the guideline, with average scores of 4.1 (without) and 4.4 (with) (p = 0.008) with a Fleiss´ Kappa of 0.44. GPT-4, when contextually enriched with a guideline, demonstrates enhanced capability in providing guideline-backed recommendations. This approach holds promise for real-time clinical decision-support, making guidelines more actionable. However, further refinements are necessary to maximize the potential of large language models (LLMs). Inherent limitations need to be addressed.

Background: Hypoxic-ischemic brain injury (HIBI) is a feared complication post-cardiac arrest (CA). The timing of brain imaging remains a topic of ongoing debate. Early computed tomography (CT) scans can reveal acute intracranial pathologies but may have limited predictive value due to delayed manifestation of HIBI-related changes. Radiomics analyses present a promising approach to identifying subtle imaging markers, potentially aiding early HIBI detection. Methods: This study retrospectively assessed post-CA patients between 2016 and 2023 who received immediate brain CTs. Patients without acute intracranial pathology on initial scans and who underwent follow-up brain CTs within 14 days post-return of spontaneous circulation (ROSC) were included. Image segmentation involved manual basalganglia segmentation and automated whole-brain segmentation. Radiomics features were calculated using Pyradiomics (v3.0.1) in 3DSlicer (v5.2.2). Feature selection involved reproducibility analysis via ICC and LASSO regression, retaining five features per segmentation method. A logistic regression model for each segmentation method underwent 5-fold cross-validation. Results were summarized with ROC analyses and average sensitivity and specificity. Results: A total of 83 patients (average age: 65 ± 13.3 years, 19 women) with CA and ROSC were included. Follow-up CT scans after 5.2 ± 2.9 days revealed brain edema in 47 patients. The model using manual segmentation achieved an average AUC of 0.76, sensitivity of 0.59, and specificity of 0.78. The automated segmentation model showed an average AUC of 0.66, sensitivity of 0.49, and specificity of 0.68. Conclusions: Radiomics, particularly focused on the basalganglia area in normal-appearing brain CTs after CA and ROSC, may enhance predictive insights for HIBI and the development of brain edema.

Background: Image quality and diagnostic accuracy in computed tomography angiography (CTA) reach their limits in imaging of below-the-knee vessels. Purpose: To evaluate whether image quality in CTA of lower limbs is further improvable by combining side-separate reconstruction with a larger matrix size and whether resulting noise can be compromised with iterative reconstruction (IR). Material and Methods: From CTA of the lower extremities of 26 patients (5 women, 21 men; mean age = 68.5 ± 10.3 years), the lower legs were reconstructed side-separately with different reconstruction algorithms and matrix sizes including filtered back projection (FBP) with a 512 × 512 matrix, FBP with a 1024 × 1024 matrix, IR (SAFIRE) with a 512 × 512 matrix, and IR (SAFIRE) with a 1024 × 1024 matrix. A total of 208 CT series were evaluated. Subjective image quality was assessed by two readers using a 5-point Likert scale. Image noise was assessed by measuring signal-to-noise and contrast-to-noise ratios. Results: Subjective image quality was rated significantly higher when using a 1024 × 1024 matrix ( P < 0.001) and could further be increased with IR. Vessel sharpness was rated significantly better with a larger matrix ( P < 0.001). Visible and measured image noise was significantly higher with a 1024 × 1024 matrix but could be reduced by using IR ( P < 0.001), even to a level below FBP with a 512 × 512 matrix while reconstructing with a larger matrix ( P < 0.001). Conclusion: Image quality, image noise, and vessel sharpness can be further improved in CTA of the lower extremities with side-separate reconstruction using a 1024 × 1024 matrix size and IR.

Objectives: The aim of this study was to determine differences in injury types and frequencies between piston-based and band-based automated chest compression devices in patients with non-traumatic out-of-hospital cardiac arrest (OHCA) at a German cardiac arrest center. Methods: This retrospective single-center study assessed resuscitation-related injuries in OHCA patients using protocol-based early whole-body CT scans at hospital admission. CT scans were reviewed independently by two reviewers blinded to the compression device used. Between May 2015 and September 2021, all patients resuscitated from non-traumatic OHCA, treated with a mechanical chest compression device, and showing stable return of spontaneous circulation (ROSC) until CT examination according to the institutional standard operating procedure for all OHCA patients were included. Patients were categorized by compression device type, and group differences were analyzed using the Chi-square test and Mann–Whitney U test. In addition, patient-level incidences of rib fracture types were calculated, and risk ratios with corresponding 95% confidence intervals were used to compare rib fracture patterns between groups. A p-value of <0.05 was considered statistically significant. Results: Among 71 patients, 32 received band-based and 39 piston-based treatment. Both groups were comparable in resuscitation duration, body constitution, and gender ratio, although the band-based group was older. Thoracic injuries predominated, with rib fractures representing the most frequent injury pattern (64/71, 90.1%). The median number of rib fractures per patient was 10 (IQR 8–12) in the band-based group and 9 (IQR 7–12) in the piston-based group. The band-based group had significantly more liver lacerations (5/32, 15.6% vs. 0/39, 0%; p = 0.01) and displaced rib fractures (117 vs. 87; p = 0.046; patient-level RR = 1.43, 95% CI 1.06–1.93). Conclusions: In this observational study of a CT-based cohort of OHCA patients with stable ROSC, the band-based device was associated with significantly higher frequencies of liver lacerations and displaced rib fractures than the piston-based device. These findings should be interpreted as hypothesis-generating and may support further evaluation of device-specific injury profiles in future studies.

Background: Acute kidney injury (AKI) is common and prognostically relevant after out-of-hospital cardiac arrest (OHCA), yet early risk prediction at admission remains limited. Objective: To assess whether radiomics features from dual-phase contrast-enhanced CT obtained at admission predict AKI after non-traumatic OHCA. Methods: This retrospective single-center study included consecutive non-traumatic OHCA patients with return of spontaneous circulation undergoing standardized admission dual-phase CT. AKI within 5 days was defined by KDIGO criteria or renal replacement therapy. Bilateral whole-kidney masks were generated using TotalSegmentator, reviewed, and corrected if necessary. Radiomics features were extracted with PyRadiomics within 3DSlicer. Features with poor reproducibility (CCC < 0.75) were excluded. Arterial-venous difference features were computed. Data were split into a training cohort and test cohort. Feature selection used mRMR followed by LASSO regression, yielding seven predictors for training of a logistic regression (LR) model, a support vector machine (SVM), and a k-nearest neighbors (kNN) model. Additionally, a limited clinical variable LR model and a combined model were evaluated. Stability was assessed by post hoc repeated resampling of the primary radiomics LR model across 100 stratified 70/30 splits, summarizing repetitions yielding 3–9 LASSO-selected features. Results: Of 383 screened patients, 155 were included, of which 47 (30.3%) developed AKI. In the test cohort, radiomics-LR showed the highest AUC (0.783), with kNN (0.778) and SVM (0.757) performing comparable. The limited clinical model performed poorly (AUC 0.549), and the combined model (AUC 0.779) did not materially improve upon radiomics alone. In the repeated resampling analysis, 29 of 100 repetitions fulfilled the predefined feature-count criterion. Among these repetitions, median test-set AUC was 0.663 (IQR, 0.623–0.699), indicating moderate but variable discriminatory performance and limited feature-selection stability. Most selected predictors were arterial-venous difference features. Conclusions: This exploratory single-center proof-of-concept study suggests that admission dual-phase CT radiomics may contain information relevant to early AKI risk stratification after OHCA, with arterial-venous difference features appearing particularly informative. However, limited stability of feature selection and model performance indicates that these findings should be considered hypothesis-generating and require methodological refinement, comparison with more comprehensive clinical models, and external validation in prospective multicenter cohorts before clinical application can be considered.