Potential of GPT-4 for Detecting Errors in Radiology Reports: Implications for Reporting Accuracy.

Background Errors in radiology reports may occur because of resident-to-attending discrepancies, speech recognition inaccuracies, and large workload. Large language models, such as GPT-4 (ChatGPT; OpenAI), may assist in generating reports. Purpose To assess effectiveness of GPT-4 in identifying common errors in radiology reports, focusing on performance, time, and cost-efficiency. Materials and Methods In this retrospective study, 200 radiology reports (radiography and cross-sectional imaging [CT and MRI]) were compiled between June 2023 and December 2023 at one institution. There were 150 errors from five common error categories (omission, insertion, spelling, side confusion, and other) intentionally inserted into 100 of the reports and used as the reference standard. Six radiologists (two senior radiologists, two attending physicians, and two residents) and GPT-4 were tasked with detecting these errors. Overall error detection performance, error detection in the five error categories, and reading time were assessed using Wald χ2 tests and paired-sample t tests. Results GPT-4 (detection rate, 82.7%;124 of 150; 95% CI: 75.8, 87.9) matched the average detection performance of radiologists independent of their experience (senior radiologists, 89.3% [134 of 150; 95% CI: 83.4, 93.3]; attending physicians, 80.0% [120 of 150; 95% CI: 72.9, 85.6]; residents, 80.0% [120 of 150; 95% CI: 72.9, 85.6]; P value range, .522-.99). One senior radiologist outperformed GPT-4 (detection rate, 94.7%; 142 of 150; 95% CI: 89.8, 97.3; P = .006). GPT-4 required less processing time per radiology report than the fastest human reader in the study (mean reading time, 3.5 seconds ± 0.5 [SD] vs 25.1 seconds ± 20.1, respectively; P < .001; Cohen d = -1.08). The use of GPT-4 resulted in lower mean correction cost per report than the most cost-efficient radiologist ($0.03 ± 0.01 vs $0.42 ± 0.41; P < .001; Cohen d = -1.12). Conclusion The radiology report error detection rate of GPT-4 was comparable with that of radiologists, potentially reducing work hours and cost. © RSNA, 2024 See also the editorial by Forman in this issue.

Advanced CT measures of coronary artery disease with intermediate stenosis in patients with severe aortic valve stenosis.

BACKGROUND: Coronary artery disease (CAD) and severe aortic valve stenosis (AS) frequently coexist. While pre-transcatheter aortic valve replacement (TAVR) computed tomography angiography (CTA) allows to rule out obstructive CAD, interpreting hemodynamic significance of intermediate stenoses is challenging. This study investigates the incremental value of CT-derived fractional flow reserve (CT-FFR), quantitative coronary plaque characteristics (e.g., stenosis degree, plaque volume, and composition), and peri-coronary adipose tissue (PCAT) density to detect hemodynamically significant lesions among those with AS and CAD. MATERIALS AND METHODS: We included patients with severe AS and intermediate coronary lesions (20-80% diameter stenosis) who underwent pre-TAVR CTA and invasive coronary angiogram (ICA) with resting full-cycle ratio (RFR) assessment between 08/16 and 04/22. CTA image analysis included assessment of CT-FFR, quantitative coronary plaque analysis, and PCAT density. Coronary lesions with RFR ≤ 0.89 indicated hemodynamic significance as reference standard. RESULTS: Overall, 87 patients (age 77.9 ± 7.4 years, 38% female) with 95 intermediate coronary artery lesions were included. CT-FFR showed good discriminatory capacity (area under receiver operator curve (AUC) = 0.89, 95% confidence interval (CI) 0.81-0.96, p < 0.001) to identify hemodynamically significant lesions, superior to anatomical assessment, plaque morphology, and PCAT density. Plaque composition and PCAT density did not differ between lesions with and without hemodynamic significance. Univariable and multivariable analyses revealed CT-FFR as the only predictor for functionally significant lesions (odds ratio 1.28 (95% CI 1.17-1.43), p < 0.001). Overall, CT-FFR ≤ 0.80 showed diagnostic accuracy, sensitivity, and specificity of 88.4% (95%CI 80.2-94.1), 78.5% (95%CI 63.2-89.7), and 96.2% (95%CI 87.0-99.5), respectively. CONCLUSION: CT-FFR was superior to CT anatomical, plaque morphology, and PCAT assessment to detect functionally significant stenoses in patients with severe AS. CLINICAL RELEVANCE STATEMENT: CT-derived fractional flow reserve in patients with severe aortic valve stenosis may be a useful tool for non-invasive hemodynamic assessment of intermediate coronary lesions, while CT anatomical, plaque morphology, and peri-coronary adipose tissue assessment have no incremental or additional benefit. These findings might help to reduce pre-transcatheter aortic valve replacement invasive coronary angiogram. KEY POINTS: • Interpreting the hemodynamic significance of intermediate coronary stenoses is challenging in pre-transcatheter aortic valve replacement CT. • CT-derived fractional flow reserve (CT-FFR) has a good discriminatory capacity in the identification of hemodynamically significant coronary lesions. • CT-derived anatomical, plaque morphology, and peri-coronary adipose tissue assessment did not improve the diagnostic capability of CT-FFR in the hemodynamic assessment of intermediate coronary stenoses.

GPT-4 for Automated Determination of Radiological Study and Protocol based on Radiology Request Forms: A Feasibility Study.

Published under a CC BY 4.0 license.

Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance.

Background Automation bias (the propensity for humans to favor suggestions from automated decision-making systems) is a known source of error in human-machine interactions, but its implications regarding artificial intelligence (AI)-aided mammography reading are unknown. Purpose To determine how automation bias can affect inexperienced, moderately experienced, and very experienced radiologists when reading mammograms with the aid of an artificial intelligence (AI) system. Materials and Methods In this prospective experiment, 27 radiologists read 50 mammograms and provided their Breast Imaging Reporting and Data System (BI-RADS) assessment assisted by a purported AI system. Mammograms were obtained between January 2017 and December 2019 and were presented in two randomized sets. The first was a training set of 10 mammograms, with the correct BI-RADS category suggested by the AI system. The second was a set of 40 mammograms in which an incorrect BI-RADS category was suggested for 12 mammograms. Reader performance, degree of bias in BI-RADS scoring, perceived accuracy of the AI system, and reader confidence in their own BI-RADS ratings were assessed using analysis of variance (ANOVA) and repeated-measures ANOVA followed by post hoc tests and Kruskal-Wallis tests followed by the Dunn post hoc test. Results The percentage of correctly rated mammograms by inexperienced (mean, 79.7% ± 11.7 [SD] vs 19.8% ± 14.0; P < .001; r = 0.93), moderately experienced (mean, 81.3% ± 10.1 vs 24.8% ± 11.6; P < .001; r = 0.96), and very experienced (mean, 82.3% ± 4.2 vs 45.5% ± 9.1; P = .003; r = 0.97) radiologists was significantly impacted by the correctness of the AI prediction of BI-RADS category. Inexperienced radiologists were significantly more likely to follow the suggestions of the purported AI when it incorrectly suggested a higher BI-RADS category than the actual ground truth compared with both moderately (mean degree of bias, 4.0 ± 1.8 vs 2.4 ± 1.5; P = .044; r = 0.46) and very (mean degree of bias, 4.0 ± 1.8 vs 1.2 ± 0.8; P = .009; r = 0.65) experienced readers. Conclusion The results show that inexperienced, moderately experienced, and very experienced radiologists reading mammograms are prone to automation bias when being supported by an AI-based system. This and other effects of human and machine interaction must be considered to ensure safe deployment and accurate diagnostic performance when combining human readers and AI. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Baltzer in this issue.

Dual-layer dual-energy CT-derived pulmonary perfusion for the differentiation of acute pulmonary embolism and chronic thromboembolic pulmonary hypertension.

OBJECTIVES: To evaluate dual-layer dual-energy computed tomography (dlDECT)-derived pulmonary perfusion maps for differentiation between acute pulmonary embolism (PE) and chronic thromboembolic pulmonary hypertension (CTEPH). METHODS: This retrospective study included 131 patients (57 patients with acute PE, 52 CTEPH, 22 controls), who underwent CT pulmonary angiography on a dlDECT. Normal and malperfused areas of lung parenchyma were semiautomatically contoured using iodine density overlay (IDO) maps. First-order histogram features of normal and malperfused lung tissue were extracted. Iodine density (ID) was normalized to the mean pulmonary artery (MPA) and the left atrium (LA). Furthermore, morphological imaging features for both acute and chronic PE, as well as the combination of histogram and morphological imaging features, were evaluated. RESULTS: In acute PE, normal perfused lung areas showed a higher mean and peak iodine uptake normalized to the MPA than in CTEPH (both p < 0.001). After normalizing mean ID in perfusion defects to the LA, patients with acute PE had a reduced average perfusion (IDmean,LA) compared to both CTEPH patients and controls (p < 0.001 for both). IDmean,LA allowed for a differentiation between acute PE and CTEPH with moderate accuracy (AUC: 0.72, sensitivity 74%, specificity 64%), resulting in a PPV and NPV for CTEPH of 64% and 70%. Combining IDmean,LA in the malperfused areas with the diameter of the MPA (MPAdia) significantly increased its ability to differentiate between acute PE and CTEPH (sole MPAdia: AUC: 0.76, 95%-CI: 0.68-0.85 vs. MPAdia + 256.3 * IDmean,LA - 40.0: AUC: 0.82, 95%-CI: 0.74-0.90, p = 0.04). CONCLUSION: dlDECT enables quantification and characterization of pulmonary perfusion patterns in acute PE and CTEPH. Although these lack precision when used as a standalone criterion, when combined with morphological CT parameters, they hold potential to enhance differentiation between the two diseases. CLINICAL RELEVANCE STATEMENT: Differentiating between acute PE and CTEPH based on morphological CT parameters is challenging, often leading to a delay in CTEPH diagnosis. By revealing distinct pulmonary perfusion patterns in both entities, dlDECT may facilitate timely diagnosis of CTEPH, ultimately improving clinical management. KEY POINTS: • Morphological imaging parameters derived from CT pulmonary angiography to distinguish between acute pulmonary embolism and chronic thromboembolic pulmonary hypertension lack diagnostic accuracy. • Dual-layer dual-energy CT reveals different pulmonary perfusion patterns between acute pulmonary embolism and chronic thromboembolic pulmonary hypertension. • The identified parameters yield potential to enable more timely identification of patients with chronic thromboembolic pulmonary hypertension.

Reduction of contrast medium for transcatheter aortic valve replacement planning using a spectral detector CT: a prospective clinical trial.

INTRODUCTION: This study investigated the use of dual-energy spectral detector computed tomography (CT) and virtual monoenergetic imaging (VMI) reconstructions in pre-interventional transcatheter aortic valve replacement (TAVR) planning. We aimed to determine the minimum required contrast medium (CM) amount to maintain diagnostic CT imaging quality for TAVR planning. METHODS: In this prospective clinical trial, TAVR candidates received a standardized dual-layer spectral detector CT protocol. The CM amount (Iohexol 350 mg iodine/mL, standardized flow rate 3 mL/s) was reduced systematically after 15 patients by 10 mL, starting at 60 mL (institutional standard). We evaluated standard, and 40- and 60-keV VMI reconstructions. For image quality, we measured signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and diameters in multiple vessel sections (i.e., aortic annulus: diameter, perimeter, area; aorta/arteries: minimal diameter). Mixed regression models (MRM), including interaction terms and clinical characteristics, were used for comparison. RESULTS: Sixty consecutive patients (mean age, 79.4 ± 7.5 years; 28 females, 46.7%) were included. In pre-TAVR CT, the CM reduction to 40 mL is possible without affecting the image quality (MRM: SNR: -1.1, p = 0.726; CNR: 0.0, p = 0.999). VMI 40-keV reconstructions showed better results than standard reconstructions with significantly higher SNR (+ 6.04, p < 0.001). Reduction to 30 mL CM resulted in a significant loss of quality (MRM: SNR: -12.9, p < 0.001; CNR: -13.9, p < 0.001), regardless of the reconstruction. Across the reconstructions, we observed no differences in the metric evaluation (p > 0.914). CONCLUSION: Among TAVR candidates undergoing pre-interventional CT at a dual-layer spectral detector system, applying 40 mL CM is sufficient to maintain diagnostic image quality. VMI 40-keV reconstructions improve the vessel attenuation and are recommended for evaluation. CLINICAL RELEVANCE STATEMENT: Contrast medium reduction to 40 mL in pre-interventional transcatheter aortic valve replacement CT using dual-energy CT maintains image quality, while 40-keV virtual monoenergetic imaging reconstructions enhance vessel attenuation. These results offer valuable recommendations for interventional transcatheter aortic valve replacement evaluation and potentially improve nephroprotection in patients with compromised renal function. KEY POINTS: • Patients undergoing transcatheter aortic valve replacement (TAVR), requiring pre-interventional CT, are often multimorbid with impaired renal function. • Using a spectral detector dual-layer CT, contrast medium reduction to 40 mL is feasible, maintaining diagnostic image quality. • The additional application of virtual monoenergetic image reconstructions with 40 keV improves vessel attenuation significantly in clinical practice.

AI support for accurate and fast radiological diagnosis of COVID-19: an international multicenter, multivendor CT study.

OBJECTIVES: Differentiation between COVID-19 and community-acquired pneumonia (CAP) in computed tomography (CT) is a task that can be performed by human radiologists and artificial intelligence (AI). The present study aims to (1) develop an AI algorithm for differentiating COVID-19 from CAP and (2) evaluate its performance. (3) Evaluate the benefit of using the AI result as assistance for radiological diagnosis and the impact on relevant parameters such as accuracy of the diagnosis, diagnostic time, and confidence. METHODS: We included n = 1591 multicenter, multivendor chest CT scans and divided them into AI training and validation datasets to develop an AI algorithm (n = 991 CT scans; n = 462 COVID-19, and n = 529 CAP) from three centers in China. An independent Chinese and German test dataset of n = 600 CT scans from six centers (COVID-19 / CAP; n = 300 each) was used to test the performance of eight blinded radiologists and the AI algorithm. A subtest dataset (180 CT scans; n = 90 each) was used to evaluate the radiologists' performance without and with AI assistance to quantify changes in diagnostic accuracy, reporting time, and diagnostic confidence. RESULTS: The diagnostic accuracy of the AI algorithm in the Chinese-German test dataset was 76.5%. Without AI assistance, the eight radiologists' diagnostic accuracy was 79.1% and increased with AI assistance to 81.5%, going along with significantly shorter decision times and higher confidence scores. CONCLUSION: This large multicenter study demonstrates that AI assistance in CT-based differentiation of COVID-19 and CAP increases radiological performance with higher accuracy and specificity, faster diagnostic time, and improved diagnostic confidence. KEY POINTS: • AI can help radiologists to get higher diagnostic accuracy, make faster decisions, and improve diagnostic confidence. • The China-German multicenter study demonstrates the advantages of a human-machine interaction using AI in clinical radiology for diagnostic differentiation between COVID-19 and CAP in CT scans.

Computed tomography-based angle measurements of the sagittal capitulum and trochlea position in relation to the humeral shaft.

The radiologic evaluation of the sagittal angulation of the distal humerus is commonly based on standard lateral radiographs. However, lateral radiographs do not allow to examine the lateral angulation of the capitulum and the trochlea, separately. Although this problem could be approached via computed tomography, there are no data available describing the difference between the angulation of the capitulum and trochlea. Therefore, we aimed to assess sagittal angles of the capitulum and trochlea in relation to the humeral shaft based on 400 CT-scans of the elbow in healthy adults. Angles were measured in sagittal planes at the capitulum center and three anatomically defined trochlea locations and were spanned between the axis of the joint component and the humerus shaft. Angles were tested for differences between measurement locations and correlation with patient characteristics (age, sex, trans-epicondylar distance). Angles increased from lateral to medial measurement locations (107.4 ± 9.6°, 167.4 ± 8.2°, 171.8 ± 7.3°, 179.1 ± 7.0°; p < 0.05). Largest angle differences were detected between the capitulum and trochlea with smallest angles measured at the capitulum. Patient characteristics did not correlate with angles (p > 0.05). Intra-rater-reliability was r = 0.79-0.86. As CT-imaging allows to distinguish between sagittal capitulum and trochlea locations, it might benefit the radiologic diagnostic of sagittal malalignments of the distal humerus at the capitulum and trochlea, separately.

Influence of aging and gadolinium exposure on T1, T2, and T2*-relaxation in healthy women with an increased risk of breast cancer with and without prior exposure to gadoterate meglumine at 3.0-T brain MR imaging.

OBJECTIVES: We examined the effects of aging and of gadolinium-based contrast agent (GBCA) exposure on MRI measurements in brain nuclei of healthy women. METHODS: This prospective, IRB-approved single-center case-control study enrolled 100 healthy participants of our high-risk screening center for hereditary breast cancer, who had received at least six doses of macrocyclic GBCA (exposed group) or were newly entering the program (GBCA-naïve group). The cutoff "at least six doses" was chosen to be able to include a sufficient number of highly exposed participants. All participants underwent unenhanced 3.0-T brain MRI including quantitative T1, T2, and R2* mapping and T1- and T2-weighted imaging. The relaxation times/signal intensities were derived from region of interest measurements in the brain nuclei performed by a radiologist and a neuroradiologist, both board certified. Statistical analysis was based on descriptive evaluations and uni-/multivariable analyses. RESULTS: The participants (exposed group: 49, control group: 51) were aged 42 ± 9 years. In a multivariable model, age had a clear impact on R2* (p < 0.001-0.012), T2 (p = 0.003-0.048), and T1 relaxation times/signal intensities (p < 0.004-0.046) for the majority of deep brain nuclei, mostly affecting the substantia nigra, globus pallidus (GP), nucleus ruber, thalamus, and dentate nucleus (DN). The effect of prior GBCA administration on T1 relaxation times was statistically significant for the DN, GP, and pons (p = 0.019-0.037). CONCLUSIONS: In a homogeneous group of young to middle-aged healthy females aging had an effect on T2 and R2* relaxation times and former GBCA applications influenced the measured T1 relaxation times. KEY POINTS: The quantitative T1, T2, and R2* relaxation times measured in women at high risk of developing breast cancer showed characteristic bandwidth for all brain nuclei examined at 3.0-T MRI. The effect of participant age had a comparatively strong impact on R2*, T2, and T1 relaxation times for the majority of brain nuclei examined. The effect of prior GBCA administrations on T1 relaxation times rates was comparatively less pronounced, yielding statistically significant results for the dentate nucleus, globus pallidus, and pons. Healthy women with and without previous GBCA-enhanced breast MRI exhibited age-related T2* and T2 relaxation alterations at 3.0 T-brain MRI. T1 relaxation alterations due to prior GBCA administration were comparatively less pronounced.

Feasibility of artificial intelligence-supported assessment of bone marrow infiltration using dual-energy computed tomography in patients with evidence of monoclonal protein - a retrospective observational study.

OBJECTIVES: To demonstrate the feasibility of an automated, non-invasive approach to estimate bone marrow (BM) infiltration of multiple myeloma (MM) by dual-energy computed tomography (DECT) after virtual non-calcium (VNCa) post-processing. METHODS: Individuals with MM and monoclonal gammopathy of unknown significance (MGUS) with concurrent DECT and BM biopsy between May 2018 and July 2020 were included in this retrospective observational study. Two pathologists and three radiologists reported BM infiltration and presence of osteolytic bone lesions, respectively. Bone mineral density (BMD) was quantified CT-based by a CE-certified software. Automated spine segmentation was implemented by a pre-trained convolutional neural network. The non-fatty portion of BM was defined as voxels > 0 HU in VNCa. For statistical assessment, multivariate regression and receiver operating characteristic (ROC) were conducted. RESULTS: Thirty-five patients (mean age 65 ± 12 years; 18 female) were evaluated. The non-fatty portion of BM significantly predicted BM infiltration after adjusting for the covariable BMD (p = 0.007, r = 0.46). A non-fatty portion of BM > 0.93% could anticipate osteolytic lesions and the clinical diagnosis of MM with an area under the ROC curve of 0.70 [0.49-0.90] and 0.71 [0.54-0.89], respectively. Our approach identified MM-patients without osteolytic lesions on conventional CT with a sensitivity and specificity of 0.63 and 0.71, respectively. CONCLUSIONS: Automated, AI-supported attenuation assessment of the spine in DECT VNCa is feasible to predict BM infiltration in MM. Further, the proposed method might allow for pre-selecting patients with higher pre-test probability of osteolytic bone lesions and support the clinical diagnosis of MM without pathognomonic lesions on conventional CT. KEY POINTS: • The retrospective study provides an automated approach for quantification of the non-fatty portion of bone marrow, based on AI-supported spine segmentation and virtual non-calcium dual-energy CT data. • An increasing non-fatty portion of bone marrow is associated with a higher infiltration determined by invasive biopsy after adjusting for bone mineral density as a control variable (p = 0.007, r = 0.46). • The non-fatty portion of bone marrow might support the clinical diagnosis of multiple myeloma when conventional CT images are negative (sensitivity 0.63, specificity 0.71).

Quality and safety of coronary computed tomography angiography at academic and non-academic sites: insights from a large European registry (ESCR MR/CT Registry).

OBJECTIVES: To compare the use of coronary computed tomography angiography (CCTA) between academic and non-academic sites across Europe over the last decade. METHODS: We analyzed a large multicenter registry (ESCR MR/CT Registry) of stable symptomatic patients who received CCTA 01/2010-01/2020 at 47 (22%) academic and 165 (78%) non-academic sites across 19 European countries. We compared image quality, radiation dose, contrast-media-related adverse events, patient characteristics, CCTA findings, and downstream testing between academic and non-academic sites. RESULTS: Among 64,317 included patients (41% female; 60 ± 13 years), academic sites accounted for most cases in 2010-2014 (52%), while non-academic sites dominated in 2015-2020 (71%). Despite less contemporary technology, non-academic sites maintained low radiation doses (4.76 [2.46-6.85] mSv) with a 30% decline of high-dose scans ( > 7 mSv) over time. Academic and non-academic sites both reported diagnostic image quality in 98% of cases and low rate of scan-related adverse events (0.4%). Academic and non-academic sites examined similar patient populations (41% females both; age: 61 ± 14 vs. 60 ± 12 years; pretest probability for obstructive CAD: low 21% vs. 23%, intermediate 73% vs. 72%, high 6% both, CAD prevalence on CCTA: 40% vs. 41%). Nevertheless, non-academic sites referred more patients to non-invasive ischemia testing (6.5% vs. 4.2%) and invasive coronary angiography/surgery (8.5% vs. 5.6%). CONCLUSIONS: Non-academic and academic sites provide safe, high-quality CCTA across Europe, essential to successfully implement the recently updated guidelines for the diagnosis and management of chronic coronary syndromes. However, despite examining similar populations with comparable CAD prevalence, non-academic sites tend to refer more patients to downstream testing. KEY POINTS: • Smaller non-academic providers increasingly use CCTA to rule out obstructive coronary artery disease. • Non-academic and academic sites provide comparably safe, high-quality CCTA across Europe. • Compared to academic sites, non-academic sites tend to refer more patients to downstream testing.

The Aortic Ductus Diverticulum-Innocent Bystander or Potential Source of Thromboembolic Stroke?

OBJECTIVE: Due to reversal blood flow in the diastolic phase, outpouchings at the aortic isthmus may carry the risk of thrombus formation and subsequent thromboembolism. The objective was to evaluate the association between aortic ductus diverticula (ADDs) and ischemic brain alterations in cerebral magnetic resonance imaging. METHODS: A retrospective analysis of 218 patients who received both a dedicated computed tomography angiography of the thoracic aorta and a brain magnetic resonance imaging was performed. Two radiologists independently reviewed all examinations for the presence of ADD as well as ischemic alterations of the brain. The association between this anatomical variant and ischemic brain alterations was evaluated by univariate and bivariate logistic regression analyses. RESULTS: ADDs were identified/present in 35 of 218 patients (16%). Ischemic brain alterations were found in 57% of patients (20/35) with an ADD and in 42% of the control group (77/183, P = 0.1). The presence of an ADD did not prove to be an independent risk factor for ischemic brain alterations after multivariate adjustment (odds ratio = 1.7, 95% confidence interval = 0.72-3.96, P = 0.225). CONCLUSIONS: In the present study, ADDs were not significantly associated with ischemic brain alterations. Therefore, ADDs seem to be an innocent bystander with respect to the pathogenesis of ischemic brain alterations.

Cerebrospinal fluid pulse wave velocity measurements: In vitro and in vivo evaluation of a novel multiband cine phase-contrast MRI sequence.

PURPOSE: Intracranial and intraspinal compliance are parameters of interest in the diagnosis and prediction of treatment outcome in patients with normal pressure hydrocephalus and other forms of communicating hydrocephalus. A noninvasive method to estimate the spinal cerebrospinal fluid (CSF) pulse wave velocity (PWV) as a measure of compliance was developed using a multiband cine phase-contrast MRI sequence and a foot-to-foot algorithm. METHODS: We used computational simulations to estimate the accuracy of the MRI acquisition and transit-time algorithm. In vitro measurements were performed to investigate the reproducibility and accuracy of the measurements under controlled conditions. In vivo measurements in 20 healthy subjects and 2 patients with normal pressure hydrocephalus were acquired to show the technical feasibility in a clinical setting. RESULTS: Simulations showed a mean deviation of the calculated CSF PWV of 3.41% ± 2.68%. In vitro results were in line with theory, showing a square-root relation between PWV and transmural pressure and a good reproducibility with SDs of repeated measurements below 5%. Mean CSF PWV over all healthy subjects was 5.83 ± 3.36 m/s. The CSF PWV measurements in the patients with normal pressure hydrocephalus were distinctly higher before CSF shunt surgery (33.80 ± 6.75 m/s and 31.31 ± 7.82 m/s), with a decrease 5 days after CSF shunt surgery (15.69 ± 3.37 m/s). CONCLUSION: This study evaluates the feasibility of CSF PWV measurements using a multiband cine phase-contrast MRI sequence. In vitro and in vivo measurements showed that this method is a potential tool for the noninvasive estimation of intraspinal compliance.

Practical applications of deep learning: classifying the most common categories of plain radiographs in a PACS using a neural network.

OBJECTIVES: The goal of the present study was to classify the most common types of plain radiographs using a neural network and to validate the network's performance on internal and external data. Such a network could help improve various radiological workflows. METHODS: All radiographs from the year 2017 (n = 71,274) acquired at our institution were retrieved from the PACS. The 30 largest categories (n = 58,219, 81.7% of all radiographs performed in 2017) were used to develop and validate a neural network (MobileNet v1.0) using transfer learning. Image categories were extracted from DICOM metadata (study and image description) and mapped to the WHO manual of diagnostic imaging. As an independent, external validation set, we used images from other institutions that had been stored in our PACS (n = 5324). RESULTS: In the internal validation, the overall accuracy of the model was 90.3% (95%CI: 89.2-91.3%), whereas, for the external validation set, the overall accuracy was 94.0% (95%CI: 93.3-94.6%). CONCLUSIONS: Using data from one single institution, we were able to classify the most common categories of radiographs with a neural network. The network showed good generalizability on the external validation set and could be used to automatically organize a PACS, preselect radiographs so that they can be routed to more specialized networks for abnormality detection or help with other parts of the radiological workflow (e.g., automated hanging protocols; check if ordered image and performed image are the same). The final AI algorithm is publicly available for evaluation and extension. KEY POINTS: • Data from one single institution can be used to train a neural network for the correct detection of the 30 most common categories of plain radiographs. • The trained model achieved a high accuracy for the majority of categories and showed good generalizability to images from other institutions. • The neural network is made publicly available and can be used to automatically organize a PACS or to preselect radiographs so that they can be routed to more specialized neural networks for abnormality detection.

Quantitative distribution of iodinated contrast media in body computed tomography: data from a large reference cohort.

OBJECTIVES: Dual-energy computed tomography allows for an accurate and reliable quantification of iodine. However, data on physiological distribution of iodine concentration (IC) is still sparse. This study aims to establish guidance for IC in abdominal organs and important anatomical landmarks using a large cohort of individuals without radiological tumor burden. METHODS: Five hundred seventy-one oncologic, portal venous phase dual-layer spectral detector CT studies of the chest and abdomen without tumor burden at time point of imaging confirmed by > 3-month follow-up were included. ROI were placed in parenchymatous organs (n = 25), lymph nodes (n = 6), and vessels (n = 3) with a minimum of two measurements per landmark. ROI were placed on conventional images and pasted to iodine maps to retrieve absolute IC. Normalization to the abdominal aorta was conducted to obtain iodine perfusion ratios. Bivariate regression analysis, t tests, and ANOVA with Tukey-Kramer post hoc test were used for statistical analysis. RESULTS: Absolute IC showed a broad scatter and varied with body mass index, between different age groups and between the sexes in parenchymatous organs, lymph nodes, and vessels (range 0.0 ± 0.0 mg/ml-6.6 ± 1.3 mg/ml). Unlike absolute IC, iodine perfusion ratios did not show dependency on body mass index; however, significant differences between the sexes and age groups persisted, showing a tendency towards decreased perfusion ratios in elderly patients (e.g., liver 18-44 years/≥ 64 years: 0.50 ± 0.11/0.43 ± 0.10, p ≤ 0.05). CONCLUSIONS: Distribution of IC obtained from a large-scale cohort is provided. As significant differences between sexes and age groups were found, this should be taken into account when obtaining quantitative iodine concentrations and applying iodine thresholds. KEY POINTS: • Absolute iodine concentration showed a broad variation and differed between body mass index, age groups, and between the sexes in parenchymatous organs, lymph nodes, and vessels. • The iodine perfusion ratios did not show dependency on body mass index while significant differences between sexes and age groups persisted. • Provided guidance values may serve as reference when aiming to differentiate healthy and abnormal tissue based on iodine perfusion ratios.

Automated detection and segmentation of thoracic lymph nodes from CT using 3D foveal fully convolutional neural networks.

BACKGROUND: In oncology, the correct determination of nodal metastatic disease is essential for patient management, as patient treatment and prognosis are closely linked to the stage of the disease. The aim of the study was to develop a tool for automatic 3D detection and segmentation of lymph nodes (LNs) in computed tomography (CT) scans of the thorax using a fully convolutional neural network based on 3D foveal patches. METHODS: The training dataset was collected from the Computed Tomography Lymph Nodes Collection of the Cancer Imaging Archive, containing 89 contrast-enhanced CT scans of the thorax. A total number of 4275 LNs was segmented semi-automatically by a radiologist, assessing the entire 3D volume of the LNs. Using this data, a fully convolutional neuronal network based on 3D foveal patches was trained with fourfold cross-validation. Testing was performed on an unseen dataset containing 15 contrast-enhanced CT scans of patients who were referred upon suspicion or for staging of bronchial carcinoma. RESULTS: The algorithm achieved a good overall performance with a total detection rate of 76.9% for enlarged LNs during fourfold cross-validation in the training dataset with 10.3 false-positives per volume and of 69.9% in the unseen testing dataset. In the training dataset a better detection rate was observed for enlarged LNs compared to smaller LNs, the detection rate for LNs with a short-axis diameter (SAD) ≥ 20 mm and SAD 5-10 mm being 91.6% and 62.2% (p < 0.001), respectively. Best detection rates were obtained for LNs located in Level 4R (83.6%) and Level 7 (80.4%). CONCLUSIONS: The proposed 3D deep learning approach achieves an overall good performance in the automatic detection and segmentation of thoracic LNs and shows reasonable generalizability, yielding the potential to facilitate detection during routine clinical work and to enable radiomics research without observer-bias.

Use and Control of Artificial Intelligence in Patients Across the Medical Workflow: Single-Center Questionnaire Study of Patient Perspectives.

BACKGROUND: Artificial intelligence (AI) is gaining increasing importance in many medical specialties, yet data on patients' opinions on the use of AI in medicine are scarce. OBJECTIVE: This study aimed to investigate patients' opinions on the use of AI in different aspects of the medical workflow and the level of control and supervision under which they would deem the application of AI in medicine acceptable. METHODS: Patients scheduled for computed tomography or magnetic resonance imaging voluntarily participated in an anonymized questionnaire between February 10, 2020, and May 24, 2020. Patient information, confidence in physicians vs AI in different clinical tasks, opinions on the control of AI, preference in cases of disagreement between AI and physicians, and acceptance of the use of AI for diagnosing and treating diseases of different severity were recorded. RESULTS: In total, 229 patients participated. Patients favored physicians over AI for all clinical tasks except for treatment planning based on current scientific evidence. In case of disagreement between physicians and AI regarding diagnosis and treatment planning, most patients preferred the physician's opinion to AI (96.2% [153/159] vs 3.8% [6/159] and 94.8% [146/154] vs 5.2% [8/154], respectively; P=.001). AI supervised by a physician was considered more acceptable than AI without physician supervision at diagnosis (confidence rating 3.90 [SD 1.20] vs 1.64 [SD 1.03], respectively; P=.001) and therapy (3.77 [SD 1.18] vs 1.57 [SD 0.96], respectively; P=.001). CONCLUSIONS: Patients favored physicians over AI in most clinical tasks and strongly preferred an application of AI with physician supervision. However, patients acknowledged that AI could help physicians integrate the most recent scientific evidence into medical care. Application of AI in medicine should be disclosed and controlled to protect patient interests and meet ethical standards.

Respiratory motion corrected 4D flow using golden radial phase encoding.

PURPOSE: To minimize respiratory motion artifacts while achieving predictable scan times with 100% scan efficiency for thoracic 4D flow MRI. METHODS: A 4D flow sequence with golden radial phase encoding (GRPE) was acquired in 9 healthy volunteers covering the heart, aorta, and venae cavae. Scan time was 15 min, and data were acquired without motion gating during acquisition. Data were retrospectively re-binned into respiratory and cardiac phases based on respiratory self-navigation and the electrocardiograph signals, respectively. Nonrigid respiratory motion fields were extracted and corrected for during the k-t SENSE reconstruction. A respiratory-motion corrected (GRPE-MOCO) and a free-breathing (GRPE-UNCORR) 4D flow dataset was reconstructed using 100% of the acquired data. For comparison, a respiratory gated Cartesian 4D flow acquisition (CART-REF) covering the aorta was acquired. Stroke volumes and peak flows were compared. Additionally, an internal flow validation based on mass conservation was performed on the GRPE-MOCO and GRPE-UNCORR. Statistically significant differences were analyzed using a paired Wilcoxon test. RESULTS: Stroke volumes and peak flows in the aorta between GRPE-MOCO and the CART-REF showed a mean difference of -1.5 ± 10.3 mL (P > 0.05) and 25.2 ± 55.9 mL/s (P > 0.05), respectively. Peak flow in the GRPE-UNCORR data was significantly different compared with CART-REF (P < 0.05). GRPE-MOCO showed higher accuracy for internal consistency analysis than GRPE-UNCORR. CONCLUSION: The proposed 4D flow sequence allows a straight-forward planning by covering the entire thorax and ensures a predictable scan time independent of cardiac cycle variations and breathing patterns.

Radiomics allows for detection of benign and malignant histopathology in patients with metastatic testicular germ cell tumors prior to post-chemotherapy retroperitoneal lymph node dissection.

OBJECTIVES: To evaluate whether a computed tomography (CT) radiomics-based machine learning classifier can predict histopathology of lymph nodes (LNs) after post-chemotherapy LN dissection (pcRPLND) in patients with metastatic non-seminomatous testicular germ cell tumors (NSTGCTs). METHODS: Eighty patients with retroperitoneal LN metastases and contrast-enhanced CT were included into this retrospective study. Resected LNs were histopathologically classified into "benign" (necrosis/fibrosis) or "malignant" (viable tumor/teratoma). On CT imaging, 204 corresponding LNs were segmented and 97 radiomic features per LN were extracted after standardized image processing. The dataset was split into training, test, and validation sets. After stepwise feature reduction based on reproducibility, variable importance, and correlation analyses, a gradient-boosted tree was trained and tuned on the selected most important features using the training and test datasets. Model validation was performed on the independent validation dataset. RESULTS: The trained machine learning classifier achieved a classification accuracy of 0.81 in the validation dataset with a misclassification of 8 of 36 benign LNs as malignant and 4 of 25 malignant LNs as benign (sensitivity 88%, specificity 72%, negative predictive value 88%). In contrast, a model containing only the LN volume resulted in a classification accuracy of 0.68 with 64% sensitivity and 68% specificity. CONCLUSIONS: CT radiomics represents an exciting new tool for improved prediction of the presence of malignant histopathology in retroperitoneal LN metastases from NSTGCTs, aiming at reducing overtreatment in this group of young patients. Thus, the presented approach should be combined with established clinical biomarkers and further validated in larger, prospective clinical trials. KEY POINTS: • Patients with metastatic non-seminomatous testicular germ cell tumors undergoing post-chemotherapy retroperitoneal lymph node dissection of residual lesions show overtreatment in up to 50%. • We assessed whether a CT radiomics-based machine learning classifier can predict histopathology of lymph nodes after post-chemotherapy lymph node dissection. • The trained machine learning classifier achieved a classification accuracy of 0.81 in the validation dataset with a sensitivity of 88% and a specificity of 78%, thus allowing for prediction of the presence of viable tumor or teratoma in retroperitoneal lymph node metastases.

Dose independent characterization of renal stones by means of dual energy computed tomography and machine learning: an ex-vivo study.

OBJECTIVES: To predict the main component of pure and mixed kidney stones using dual-energy computed tomography and machine learning. METHODS: 200 kidney stones with a known composition as determined by infrared spectroscopy were examined using a non-anthropomorphic phantom on a spectral detector computed tomography scanner. Stones were of either pure (monocrystalline, n = 116) or compound (dicrystalline, n = 84) composition. Image acquisition was repeated twice using both, normal and low-dose protocols, respectively (ND/LD). Conventional images and low and high keV virtual monoenergetic images were reconstructed. Stones were semi-automatically segmented. A shallow neural network was trained using data from ND1 acquisition split into training (70%), testing (15%) and validation-datasets (15%). Performance for ND2 and both LD acquisitions was tested. Accuracy on a per-voxel and a per-stone basis was calculated. RESULTS: Main components were: Whewellite (n = 80), weddellite (n = 21), Ca-phosphate (n = 39), cysteine (n = 20), struvite (n = 13), uric acid (n = 18) and xanthine stones (n = 9). Stone size ranged from 3 to 18 mm. Overall accuracy for predicting the main component on a per-voxel basis attained by ND testing dataset was 91.1%. On independently tested acquisitions, accuracy was 87.1-90.4%. CONCLUSIONS: Even in compound stones, the main component can be reliably determined using dual energy CT and machine learning, irrespective of dose protocol. KEY POINTS: • Spectral Detector Dual Energy CT and Machine Learning allow for an accurate prediction of stone composition. • Ex-vivo study demonstrates the dose independent assessment of pure and compound stones. • Lowest accuracy is reported for compound stones with struvite as main component.