ABSTRACT
OBJECTIVES: Diagnostic accuracy of artificial intelligence (AI) pneumothorax (PTX) detection in chest radiographs (CXR) is limited by the noisy annotation quality of public training data and confounding thoracic tubes (TT). We hypothesize that in-image annotations of the dehiscent visceral pleura for algorithm training boosts algorithm's performance and suppresses confounders. METHODS: Our single-center evaluation cohort of 3062 supine CXRs includes 760 PTX-positive cases with radiological annotations of PTX size and inserted TTs. Three step-by-step improved algorithms (differing in algorithm architecture, training data from public datasets/clinical sites, and in-image annotations included in algorithm training) were characterized by area under the receiver operating characteristics (AUROC) in detailed subgroup analyses and referenced to the well-established "CheXNet" algorithm. RESULTS: Performances of established algorithms exclusively trained on publicly available data without in-image annotations are limited to AUROCs of 0.778 and strongly biased towards TTs that can completely eliminate algorithm's discriminative power in individual subgroups. Contrarily, our final "algorithm 2" which was trained on a lower number of images but additionally with in-image annotations of the dehiscent pleura achieved an overall AUROC of 0.877 for unilateral PTX detection with a significantly reduced TT-related confounding bias. CONCLUSIONS: We demonstrated strong limitations of an established PTX-detecting AI algorithm that can be significantly reduced by designing an AI system capable of learning to both classify and localize PTX. Our results are aimed at drawing attention to the necessity of high-quality in-image localization in training data to reduce the risks of unintentionally biasing the training process of pathology-detecting AI algorithms. KEY POINTS: ⢠Established pneumothorax-detecting artificial intelligence algorithms trained on public training data are strongly limited and biased by confounding thoracic tubes. ⢠We used high-quality in-image annotated training data to effectively boost algorithm performance and suppress the impact of confounding thoracic tubes. ⢠Based on our results, we hypothesize that even hidden confounders might be effectively addressed by in-image annotations of pathology-related image features.
Subject(s)
Artificial Intelligence , Pneumothorax , Algorithms , Data Curation , Humans , Pneumothorax/diagnostic imaging , Radiography , Radiography, ThoracicABSTRACT
BACKGROUND: Characteristics of COVID-19 patients have mainly been reported within confirmed COVID-19 cohorts. By analyzing patients with respiratory infections in the emergency department during the first pandemic wave, we aim to assess differences in the characteristics of COVID-19 vs. Non-COVID-19 patients. This is particularly important regarding the second COVID-19 wave and the approaching influenza season. METHODS: We prospectively included 219 patients with suspected COVID-19 who received radiological imaging and RT-PCR for SARS-CoV-2. Demographic, clinical and laboratory parameters as well as RT-PCR results were used for subgroup analysis. Imaging data were reassessed using the following scoring system: 0 - not typical, 1 - possible, 2 - highly suspicious for COVID-19. RESULTS: COVID-19 was diagnosed in 72 (32,9%) patients. In three of them (4,2%) the initial RT-PCR was negative while initial CT scan revealed pneumonic findings. 111 (50,7%) patients, 61 of them (55,0%) COVID-19 positive, had evidence of pneumonia. Patients with COVID-19 pneumonia showed higher body temperature (37,7 ± 0,1 vs. 37,1 ± 0,1 °C; p = 0.0001) and LDH values (386,3 ± 27,1 vs. 310,4 ± 17,5 U/l; p = 0.012) as well as lower leukocytes (7,6 ± 0,5 vs. 10,1 ± 0,6G/l; p = 0.0003) than patients with other pneumonia. Among abnormal CT findings in COVID-19 patients, 57 (93,4%) were evaluated as highly suspicious or possible for COVID-19. In patients with negative RT-PCR and pneumonia, another third was evaluated as highly suspicious or possible for COVID-19 (14 out of 50; 28,0%). The sensitivity in the detection of patients requiring isolation was higher with initial chest CT than with initial RT-PCR (90,4% vs. 79,5%). CONCLUSIONS: COVID-19 patients show typical clinical, laboratory and imaging parameters which enable a sensitive detection of patients who demand isolation measures due to COVID-19.
Subject(s)
COVID-19/diagnosis , COVID-19/physiopathology , Respiratory Tract Infections/diagnosis , Respiratory Tract Infections/physiopathology , Adult , Aged , Aged, 80 and over , COVID-19/epidemiology , COVID-19 Nucleic Acid Testing , Emergency Service, Hospital , Female , Germany/epidemiology , Hospitalization , Humans , Lung/diagnostic imaging , Male , Middle Aged , Pandemics , Prospective Studies , Respiratory Tract Infections/epidemiology , SARS-CoV-2 , Tomography, X-Ray Computed , Young AdultABSTRACT
BACKGROUND: Macrophages engulf particulate contrast media, which is pivotal for biomedical imaging. PURPOSE: To introduce a macrophage ablation animal model by showing its power to manipulate the kinetics of imaging probes. MATERIAL AND METHODS: The kinetics of a particulate computed tomography (CT) contrast media was compared in macrophage ablative mice and normal mice. Liposomes (size 220 µg), loaded with clodronate, were injected into the peritoneum of three C57BL/6 mice. On the third day, 200 µL of the particulate agent ExiTron nano 6000 were injected into three macrophage-ablative mice and three control mice. CT scans were acquired before and 3 min, 1 h, 6 h, and 24 h after the ExiTron application. The animals were sacrificed, and their spleens and livers removed. Relative CT values (CTV) were measured and analyzed. RESULTS: Liver and spleen enhancement of treated mice and controls were increasing over time. The median peak values were different with 225 CTV for treated mice and 582 CTV for controls in the liver (P = 0.032) and 431 CTV for treated and 974 CTV in controls in the spleen (P = 0.016). CONCLUSION: Macrophage ablation leads to a decrease of enhancement in organs containing high numbers of macrophages, but only marginal changes in macrophage-poor organs. Macrophage ablation can influence the phagocytic activity and thus opens new potentials to investigate and manipulate the uptake of imaging probes.
Subject(s)
Ablation Techniques , Clodronic Acid/administration & dosage , Contrast Media/pharmacokinetics , Liver/metabolism , Macrophages/drug effects , Spleen/metabolism , Animals , Female , Liposomes , Liver/diagnostic imaging , Macrophages/metabolism , Macrophages/pathology , Mice , Mice, Inbred C57BL , Models, Animal , Mononuclear Phagocyte System , Spleen/diagnostic imaging , Tomography, X-Ray ComputedABSTRACT
OBJECTIVES: Interpretation of lung opacities in ICU supine chest radiographs remains challenging. We evaluated a prototype artificial intelligence algorithm to classify basal lung opacities according to underlying pathologies. DESIGN: Retrospective study. The deep neural network was trained on two publicly available datasets including 297,541 images of 86,876 patients. PATIENTS: One hundred sixty-six patients received both supine chest radiograph and CT scans (reference standard) within 90 minutes without any intervention in between. MEASUREMENTS AND MAIN RESULTS: Algorithm accuracy was referenced to board-certified radiologists who evaluated supine chest radiographs according to side-separate reading scores for pneumonia and effusion (0 = absent, 1 = possible, and 2 = highly suspected). Radiologists were blinded to the supine chest radiograph findings during CT interpretation. Performances of radiologists and the artificial intelligence algorithm were quantified by receiver-operating characteristic curve analysis. Diagnostic metrics (sensitivity, specificity, positive predictive value, negative predictive value, and accuracy) were calculated based on different receiver-operating characteristic operating points. Regarding pneumonia detection, radiologists achieved a maximum diagnostic accuracy of up to 0.87 (95% CI, 0.78-0.93) when considering only the supine chest radiograph reading score 2 as positive for pneumonia. Radiologist's maximum sensitivity up to 0.87 (95% CI, 0.76-0.94) was achieved by additionally rating the supine chest radiograph reading score 1 as positive for pneumonia and taking previous examinations into account. Radiologic assessment essentially achieved nonsignificantly higher results compared with the artificial intelligence algorithm: artificial intelligence-area under the receiver-operating characteristic curve of 0.737 (0.659-0.815) versus radiologist's area under the receiver-operating characteristic curve of 0.779 (0.723-0.836), diagnostic metrics of receiver-operating characteristic operating points did not significantly differ. Regarding the detection of pleural effusions, there was no significant performance difference between radiologist's and artificial intelligence algorithm: artificial intelligence-area under the receiver-operating characteristic curve of 0.740 (0.662-0.817) versus radiologist's area under the receiver-operating characteristic curve of 0.698 (0.646-0.749) with similar diagnostic metrics for receiver-operating characteristic operating points. CONCLUSIONS: Considering the minor level of performance differences between the algorithm and radiologists, we regard artificial intelligence as a promising clinical decision support tool for supine chest radiograph examinations in the clinical routine with high potential to reduce the number of missed findings in an artificial intelligence-assisted reading setting.
Subject(s)
Artificial Intelligence , Critical Illness/epidemiology , Image Interpretation, Computer-Assisted , Lung Diseases/diagnostic imaging , Radiography, Thoracic , Algorithms , Female , Humans , Lung Diseases/diagnosis , Male , Middle Aged , Radiologists/standards , Radiologists/statistics & numerical data , Reproducibility of Results , Retrospective Studies , Supine Position , Tomography, X-Ray ComputedABSTRACT
Purpose To determine the impact of patient age on the cost-effectiveness of endovascular therapy (EVT) in addition to standard care (SC) in large-vessel-occlusion stroke for patients aged 50 to 100 years in the United States. Materials and Methods A decision-analytic Markov model was used to estimate direct and indirect lifetime costs and quality-adjusted life years (QALYs). Age-dependent input parameters were obtained from the literature. Deterministic and probabilistic sensitivity analysis for age at index stroke were used. The willingness-to-pay (WTP) was set to thresholds of $50 000, $100 000, and $150 000 per QALY. The study applied a U.S. setting for health care and societal perspectives. Incremental costs and effectiveness were derived from deterministic and probabilistic sensitivity analysis. Acceptability rates at different WTP thresholds were determined. Results EVT+SC was the dominant strategy in patients aged 50 to 79 years. The highest incremental effectiveness (2.61 QALYs) and cost-savings (health care perspective, $99 555; societal perspective, $146 385) were obtained in 50-year-old patients. In octogenarians (80-89 years), EVT+SC led to incremental QALYs at incremental costs with acceptability rates of more than 85%, more than 99%, and more than 99% at a WTP of $50 000, $100 000, and $150 000 per QALY, respectively. In nonagenarians (90-99 years), acceptability rates at a WTP of $50 000 per QALY dropped but stayed higher than 85% and higher than 95% at thresholds of $100 000 and $150 000 per QALY. Conclusion Using contemporary willingness-to-pay thresholds in the United States, endovascular therapy in addition to standard care reduces lifetime costs for patients up to 79 years of age and is cost-effective for patients aged 80 to 100 years.
Subject(s)
Cost-Benefit Analysis/economics , Endovascular Procedures/economics , Endovascular Procedures/methods , Stroke/economics , Stroke/therapy , Brain Ischemia/complications , Brain Ischemia/economics , Brain Ischemia/therapy , Cost-Benefit Analysis/statistics & numerical data , Humans , Stroke/complicationsABSTRACT
PURPOSE: To compare free text (FTR) and structured reports (SR) of videofluoroscopic swallowing studies (VFSS) and evaluate satisfaction of referring otolaryngologists and speech therapists. MATERIALS AND METHODS: Both standard FTR and SR of 26 patients with VFSS were acquired. A dedicated template focusing on oropharyngeal phases was created for SR using online software with clickable decision-trees and concomitant generation of semantically structured reports. All reports were evaluated regarding overall quality and content, information extraction and clinical decision support (10-point Likert scale (0 = I completely disagree, 10 = I completely agree)). RESULTS: Two otorhinolaryngologists and two speech therapists evaluated FTR and SR. SR received better ratings than FTR in all items. SR were perceived to contain more details on the swallowing phases (median rating: 10 vs. 5; P < 0.001), penetration and aspiration (10 vs. 5; P < 0.001) and facilitated information extraction compared to FTR (10 vs. 4; P < 0.001). Overall quality was rated significantly higher in SR than FTR (P < 0.001). CONCLUSION: SR of VFSS provide more detailed information and facilitate information extraction. SR better assist in clinical decision-making, might enhance the quality of the report and, thus, are recommended for the evaluation of VFSS. KEY POINTS: ⢠Structured reports on videofluoroscopic exams of deglutition lead to improved report quality. ⢠Information extraction is facilitated when using structured reports based on decision trees. ⢠Template-based reports add more value to clinical decision-making than free text reports. ⢠Structured reports receive better ratings by speech therapists and otolaryngologists. ⢠Structured reports on videofluoroscopic exams may improve the comparability between exams.
Subject(s)
Deglutition Disorders/diagnostic imaging , Medical Records/statistics & numerical data , Quality Improvement/statistics & numerical data , Video Recording , Aged , Esophagus/diagnostic imaging , Female , Fluoroscopy/methods , Humans , Male , Medical Records/standards , Middle Aged , Pharynx/diagnostic imaging , Retrospective StudiesABSTRACT
OBJECTIVES: The aim was to evaluate the effect of structured reporting of computed tomography angiography (CTA) runoff studies on clarity, completeness, clinical relevance, usefulness of the radiology reports, further testing, and therapy in patients with known or suspected peripheral arterial disease. METHODS: Conventional reports (CRs) and structured reports (SRs) were generated for 52 patients who had been examined with a CTA runoff examination of the lower extremities. The sample size was based on power calculations with a power of 95% and a significance level of .007 (adjusted for multiple testing). CRs were dictated in a free text form; SRs contained a consistent ordering of observations with standardised subheadings. CRs were compared with SRs. Two vascular medicine specialists and two vascular surgeons rated the reports regarding their satisfaction with clarity, completeness, clinical relevance, and usefulness as well as overall satisfaction. Additionally, they made hypothetical decisions on further testing and therapy. Median ratings were compared using the Wilcoxon signed rank test and generalised linear mixed effects models. RESULTS: SRs received higher ratings for satisfaction with clarity (median rating 9.0 vs. 7.0, p < .0001) and completeness (median rating 9.0 vs. 7.5, p < .0001) and were judged to be of greater clinical relevance (median rating 9.0 vs. 8.0, p < .0001) and usefulness (median rating 9.0 vs. 8.0, p < .0001). Overall satisfaction was also higher for SRs (median rating 9.0 vs. 7.0, p < .0001) than CRs. There were no significant differences in further testing or therapy. CONCLUSION: Referring clinicians perceive SRs of CTA runoff examinations of the lower extremities as offering superior clarity, completeness, clinical relevance, and usefulness than CRs. Structured reporting does not appear to alter further testing or therapy in patients with known or suspected peripheral arterial disease.
Subject(s)
Computed Tomography Angiography , Lower Extremity/blood supply , Peripheral Arterial Disease/diagnosis , Aged , Computed Tomography Angiography/methods , Computed Tomography Angiography/standards , Data Accuracy , Female , Humans , Male , Predictive Value of Tests , Reproducibility of ResultsABSTRACT
BACKGROUND: To analyse structured and free text reports of shoulder X-ray examinations evaluating the quality of reports and potential contributions to clinical decision-making. METHODS: We acquired both standard free text and structured reports of 31 patients with a painful shoulder without history of previous trauma who received X-ray exams. A template was created for the structured report based on the template ID 0000154 (Shoulder X-ray) from radreport.org using online software with clickable decision trees with concomitant generation of structured semantic reports. All reports were evaluated regarding overall quality and key features: content, information extraction and clinical relevance. RESULTS: Two experienced orthopaedic surgeons reviewed and rated structured and free text reports of 31 patients independently. The structured reports achieved significantly higher median ratings in all key features evaluated (P < 0.001), including facilitation of information extraction (P < 0.001) and better contribution to subsequent clinical decision-making (P < 0.001). The overall quality of structured reports was significantly higher than in free text report (P < 0.001). CONCLUSIONS: A comprehensive structured template may be a useful tool to assist in clinical decision-making and is, thus, recommended for the reporting of degenerative changes regarding X-ray examinations of the shoulder.
Subject(s)
Medical Records/classification , Medical Records/standards , Shoulder Pain/diagnostic imaging , Clinical Decision-Making , Female , Humans , Interdisciplinary Communication , Internet , Male , Radiography , Research Report/standards , Retrospective Studies , SoftwareABSTRACT
BACKGROUND: Intracranial arterial calcifications (ICAC) are often detected on unenhanced CT of patients with an age > 60. However, association with the subsequent occurrence of major adverse cardiovascular events (MACE) has not yet been evaluated. PURPOSE: This study aimed at evaluating the association of ICAC with subsequent MACE and overall mortality. METHODS: In this retrospective, IRB approved study, we included 175 consecutive patients (89 males, mean age 78.3 ± 8.5 years) of age > 60 years who underwent an unenhanced CT of the head due to minor trauma or neurological disorders. Presence of ICAC was determined in seven intracranial arteries using a semi-quantitative scale, which resulted in the calcified plaque score (CPS). Clinical follow-up information was obtained by questionnaires and telephone interviews. MACE was defined as myocardial infarction or revascularization, stroke or death due to cardiovascular event. RESULTS: Mean follow-up time was 39.8 ± 7.8 months, resulting in 579.7 patient-years of follow-up. Overall, 36 MACE occurred during follow-up (annual event rate = 6.2%/year). Mean CPS was significantly higher in subjects with MACE during follow-up compared to subjects without MACE (p < 0.01). In 15 patients CPS was 0; in none of these patients MACE was registered. Kaplan-Meier-analysis revealed that patients with a low plaque burden (CPS < 5) had a significant longer MACE-free and overall survival than patients with a high plaque burden (CPS ≥ 5) (p < 0.01). CONCLUSION: Patients with ICAC have an increased risk for future cardio- or cerebrovascular events. Therefore, ICAC might be a prognostic factor to determine the risk for these events in older patients.
Subject(s)
Cardiovascular Diseases/etiology , Cerebrovascular Disorders/complications , Cerebrovascular Disorders/diagnostic imaging , Tomography, X-Ray Computed/methods , Vascular Calcification/complications , Vascular Calcification/diagnostic imaging , Aged , Aged, 80 and over , Cardiovascular Diseases/mortality , Female , Humans , Interviews as Topic , Male , Middle Aged , Predictive Value of Tests , Prognosis , Retrospective Studies , Surveys and Questionnaires , Survival Rate , Vascular Calcification/mortalityABSTRACT
BACKGROUND AND PURPOSE: Malignant cerebellar edema (MCE) is a life-threatening complication of acute ischemic stroke that requires timely diagnosis and management. Aim of this study was to identify imaging predictors in initial multiparametric computed tomography (CT), including whole-brain CT perfusion (WB-CTP). METHODS: We consecutively selected all subjects with cerebellar ischemic WB-CTP deficits and follow-up-confirmed cerebellar infarction from an initial cohort of 2635 patients who had undergone multiparametric CT because of suspected stroke. Follow-up imaging was assessed for the presence of MCE, measured using an established 10-point scale, of which scores ≥4 are considered malignant. Posterior circulation-Acute Stroke Prognosis Early CT Score (pc-ASPECTS) was determined to assess ischemic changes on noncontrast CT, CT angiography (CTA), and parametric WB-CTP maps (cerebellar blood flow [CBF]; cerebellar blood volume; mean transit time; time to drain). Fisher's exact tests, Mann-Whitney U tests, and receiver operating characteristics analyses were performed for statistical analyses. RESULTS: Out of a total of 51 patients who matched the inclusion criteria, 42 patients (82.4%) were categorized as MCE- and 9 (17.6%) as MCE+. MCE+ patients had larger CBF, cerebellar blood volume, mean transit time, and time to drain deficit volumes (all with P<0.001) and showed significantly lower median pc-ASPECTS assessed using WB-CTP (CBF, cerebellar blood volume, mean transit time, time to drain; all with P<0.001) compared with MCE- patients, while median pc-ASPECTS on noncontrast CT and CTA was not significantly different (both P>0.05). Receiver operating characteristics analyses yielded the largest area under the curve values for the prediction of MCE development for CBF (0.979) and cerebellar blood volume deficit volumes (0.956) and pc-ASPECTS on CBF (0.935), whereas pc-ASPECTS on noncontrast CT (0.648) and CTA (0.684) had less diagnostic value. The optimal cutoff value for CBF deficit volume was 22 mL, yielding 100% sensitivity and 90% specificity for MCE classification. CONCLUSIONS: WB-CTP provides added diagnostic value for the early identification of patients at risk for MCE development in acute cerebellar stroke.
Subject(s)
Brain Edema/diagnostic imaging , Brain Infarction/diagnostic imaging , Cerebellar Diseases/diagnostic imaging , Cerebellum/blood supply , Aged , Aged, 80 and over , Brain Edema/etiology , Brain Infarction/complications , Cerebellar Diseases/complications , Female , Humans , Male , Middle Aged , Perfusion Imaging , ROC Curve , Stroke/diagnostic imaging , Tomography, X-Ray ComputedABSTRACT
Clinical imaging uses a variety of medical imaging techniques to diagnose and monitor diseases, injuries and other health conditions. These include Xray images, computed tomography (CT), magnetic resonance imaging (MRI) and ultrasound. These procedures are used to make accurate diagnoses and plan the best possible treatment for patients. Forensic imaging, in contrast, is used in both living and deceased persons in the context of criminal investigations. Postmortem forensic imaging techniques, such as postmortem CT (PMCT) and postmortem CT angiography (PMCTA), include some of the same procedures used in clinical imaging. An important difference between clinical and forensic imaging is the purpose and context in which the imaging studies are used. In addition, radiological procedures, such as angiography, need to be adapted and modified in the post-mortem setting. From a legal perspective clinical and forensic imaging must strictly adhere to privacy and procedural guidelines. Forensic images often need to be admissible as evidence in court, which places specific requirements on the quality, authenticity and documentation of images. In the case of living individuals, there must be a valid indication and consent from the patient. Consent must also fundamentally be obtained for post-mortem examinations, e.g. from the public prosecutor's office.
ABSTRACT
PURPOSE: The aim of this study was to develop an algorithm to automatically extract annotations from German thoracic radiology reports to train deep learning-based chest X-ray classification models. MATERIALS AND METHODS: An automatic label extraction model for German thoracic radiology reports was designed based on the CheXpert architecture. The algorithm can extract labels for twelve common chest pathologies, the presence of support devices, and "no finding". For iterative improvements and to generate a ground truth, a web-based multi-reader annotation interface was created. With the proposed annotation interface, a radiologist annotated 1086 retrospectively collected radiology reports from 2020-2021 (data set 1). The effect of automatically extracted labels on chest radiograph classification performance was evaluated on an additional, in-house pneumothorax data set (data set 2), containing 6434 chest radiographs with corresponding reports, by comparing a DenseNet-121 model trained on extracted labels from the associated reports, image-based pneumothorax labels, and publicly available data, respectively. RESULTS: Comparing automated to manual labeling on data set 1: "mention extraction" class-wise F1 scores ranged from 0.8 to 0.995, the "negation detection" F1 scores from 0.624 to 0.981, and F1 scores for "uncertainty detection" from 0.353 to 0.725. Extracted pneumothorax labels on data set 2 had a sensitivity of 0.997 [95â% CI: 0.994, 0.999] and specificity of 0.991 [95â% CI: 0.988, 0.994]. The model trained on publicly available data achieved an area under the receiver operating curve (AUC) for pneumothorax classification of 0.728 [95â% CI: 0.694, 0.760], while the models trained on automatically extracted labels and on manual annotations achieved values of 0.858 [95â% CI: 0.832, 0.882] and 0.934 [95â% CI: 0.918, 0.949], respectively. CONCLUSION: Automatic label extraction from German thoracic radiology reports is a promising substitute for manual labeling. By reducing the time required for data annotation, larger training data sets can be created, resulting in improved overall modeling performance. Our results demonstrated that a pneumothorax classifier trained on automatically extracted labels strongly outperformed the model trained on publicly available data, without the need for additional annotation time and performed competitively compared to manually labeled data. KEY POINTS: · An algorithm for automatic German thoracic radiology report annotation was developed.. · Automatic label extraction is a promising substitute for manual labeling.. · The classifier trained on extracted labels outperformed the model trained on publicly available data.. ZITIERWEISE: · Wollek A, Hyska S, Sedlmeyr T etâal. German CheXpert Chest X-ray Radiology Report Labeler. Fortschr Röntgenstr 2024; 196: 956â-â965.
Subject(s)
Algorithms , Radiography, Thoracic , Radiography, Thoracic/methods , Humans , Germany , Retrospective Studies , Pneumothorax/diagnostic imaging , Neural Networks, ComputerABSTRACT
The aim of this study was to explore the potential of weak supervision in a deep learning-based label prediction model. The goal was to use this model to extract labels from German free-text thoracic radiology reports on chest X-ray images and for training chest X-ray classification models.The proposed label extraction model for German thoracic radiology reports uses a German BERT encoder as a backbone and classifies a report based on the CheXpert labels. For investigating the efficient use of manually annotated data, the model was trained using manual annotations, weak rule-based labels, and both. Rule-based labels were extracted from 66071 retrospectively collected radiology reports from 2017-2021 (DS 0), and 1091 reports from 2020-2021 (DS 1) were manually labeled according to the CheXpert classes. Label extraction performance was evaluated with respect to mention extraction, negation detection, and uncertainty detection by measuring F1 scores. The influence of the label extraction method on chest X-ray classification was evaluated on a pneumothorax data set (DS 2) containing 6434 chest radiographs with associated reports and expert diagnoses of pneumothorax. For this, DenseNet-121 models trained on manual annotations, rule-based and deep learning-based label predictions, and publicly available data were compared.The proposed deep learning-based labeler (DL) performed on average considerably stronger than the rule-based labeler (RB) for all three tasks on DS 1 with F1 scores of 0.938 vs. 0.844 for mention extraction, 0.891 vs. 0.821 for negation detection, and 0.624 vs. 0.518 for uncertainty detection. Pre-training on DS 0 and fine-tuning on DS 1 performed better than only training on either DS 0 or DS 1. Chest X-ray pneumothorax classification results (DS 2) were highest when trained with DL labels with an area under the receiver operating curve (AUC) of 0.939 compared to RB labels with an AUC of 0.858. Training with manual labels performed slightly worse than training with DL labels with an AUC of 0.934. In contrast, training with a public data set resulted in an AUC of 0.720.Our results show that leveraging a rule-based report labeler for weak supervision leads to improved labeling performance. The pneumothorax classification results demonstrate that our proposed deep learning-based labeler can serve as a substitute for manual labeling requiring only 1000 manually annotated reports for training. · The proposed deep learning-based label extraction model for German thoracic radiology reports performs better than the rule-based model.. · Training with limited supervision outperformed training with a small manually labeled data set.. · Using predicted labels for pneumothorax classification from chest radiographs performed equally to using manual annotations.. Wollek A, Haitzer P, Sedlmeyr T et al. Language modelbased labeling of German thoracic radiology reports. Fortschr Röntgenstr 2024; DOI 10.1055/a-2287-5054.
ABSTRACT
BACKGROUND: Chest radiographs (CXRs) are still of crucial importance in primary diagnostics, but their interpretation poses difficulties at times. RESEARCH QUESTION: Can a convolutional neural network-based artificial intelligence (AI) system that interprets CXRs add value in an emergency unit setting? STUDY DESIGN AND METHODS: A total of 563 CXRs acquired in the emergency unit of a major university hospital were retrospectively assessed twice by three board-certified radiologists, three radiology residents, and three emergency unit-experienced nonradiology residents (NRRs). They used a two-step reading process: (1) without AI support; and (2) with AI support providing additional images with AI overlays. Suspicion of four suspected pathologies (pleural effusion, pneumothorax, consolidations suspicious for pneumonia, and nodules) was reported on a five-point confidence scale. Confidence scores of the board-certified radiologists were converted into four binary reference standards of different sensitivities. Performance by radiology residents and NRRs without AI support/with AI support were statistically compared by using receiver-operating characteristics (ROCs), Youden statistics, and operating point metrics derived from fitted ROC curves. RESULTS: NRRs could significantly improve performance, sensitivity, and accuracy with AI support in all four pathologies tested. In the most sensitive reference standard (reference standard IV), NRR consensus improved the area under the ROC curve (mean, 95% CI) in the detection of the time-critical pathology pneumothorax from 0.846 (0.785-0.907) without AI support to 0.974 (0.947-1.000) with AI support (P < .001), which represented a gain of 30% in sensitivity and 2% in accuracy (while maintaining an optimized specificity). The most pronounced effect was observed in nodule detection, with NRR with AI support improving sensitivity by 53% and accuracy by 7% (area under the ROC curve without AI support, 0.723 [0.661-0.785]; with AI support, 0.890 [0.848-0.931]; P < .001). Radiology residents had smaller, mostly nonsignificant gains in performance, sensitivity, and accuracy with AI support. INTERPRETATION: We found that in an emergency unit setting without 24/7 radiology coverage, the presented AI solution features an excellent clinical support tool to nonradiologists, similar to a second reader, and allows for a more accurate primary diagnosis and thus earlier therapy initiation.
Subject(s)
Artificial Intelligence , Emergency Service, Hospital , Radiography, Thoracic , Humans , Radiography, Thoracic/methods , Retrospective Studies , Male , Female , Clinical Competence , Middle Aged , ROC Curve , Adult , AgedABSTRACT
Purpose: To investigate the chest radiograph classification performance of vision transformers (ViTs) and interpretability of attention-based saliency maps, using the example of pneumothorax classification. Materials and Methods: In this retrospective study, ViTs were fine-tuned for lung disease classification using four public datasets: CheXpert, Chest X-Ray 14, MIMIC CXR, and VinBigData. Saliency maps were generated using transformer multimodal explainability and gradient-weighted class activation mapping (GradCAM). Classification performance was evaluated on the Chest X-Ray 14, VinBigData, and Society for Imaging Informatics in Medicine-American College of Radiology (SIIM-ACR) Pneumothorax Segmentation datasets using the area under the receiver operating characteristic curve (AUC) analysis and compared with convolutional neural networks (CNNs). The explainability methods were evaluated with positive and negative perturbation, sensitivity-n, effective heat ratio, intra-architecture repeatability, and interarchitecture reproducibility. In the user study, three radiologists classified 160 chest radiographs with and without saliency maps for pneumothorax and rated their usefulness. Results: ViTs had comparable chest radiograph classification AUCs compared with state-of-the-art CNNs: 0.95 (95% CI: 0.94, 0.95) versus 0.83 (95%, CI 0.83, 0.84) on Chest X-Ray 14, 0.84 (95% CI: 0.77, 0.91) versus 0.83 (95% CI: 0.76, 0.90) on VinBigData, and 0.85 (95% CI: 0.85, 0.86) versus 0.87 (95% CI: 0.87, 0.88) on SIIM-ACR. Both saliency map methods unveiled a strong bias toward pneumothorax tubes in the models. Radiologists found 47% of the attention-based and 39% of the GradCAM saliency maps useful. The attention-based methods outperformed GradCAM on all metrics. Conclusion: ViTs performed similarly to CNNs in chest radiograph classification, and their attention-based saliency maps were more useful to radiologists and outperformed GradCAM.Keywords: Conventional Radiography, Thorax, Diagnosis, Supervised Learning, Convolutional Neural Network (CNN) Online supplemental material is available for this article. © RSNA, 2023.
ABSTRACT
Artificial intelligence (AI) algorithms evaluating [supine] chest radiographs ([S]CXRs) have remarkably increased in number recently. Since training and validation are often performed on subsets of the same overall dataset, external validation is mandatory to reproduce results and reveal potential training errors. We applied a multicohort benchmarking to the publicly accessible (S)CXR analyzing AI algorithm CheXNet, comprising three clinically relevant study cohorts which differ in patient positioning ([S]CXRs), the applied reference standards (CT-/[S]CXR-based) and the possibility to also compare algorithm classification with different medical experts' reading performance. The study cohorts include [1] a cohort, characterized by 563 CXRs acquired in the emergency unit that were evaluated by 9 readers (radiologists and non-radiologists) in terms of 4 common pathologies, [2] a collection of 6,248 SCXRs annotated by radiologists in terms of pneumothorax presence, its size and presence of inserted thoracic tube material which allowed for subgroup and confounding bias analysis and [3] a cohort consisting of 166 patients with SCXRs that were evaluated by radiologists for underlying causes of basal lung opacities, all of those cases having been correlated to a timely acquired computed tomography scan (SCXR and CT within < 90 min). CheXNet non-significantly exceeded the radiology resident (RR) consensus in the detection of suspicious lung nodules (cohort [1], AUC AI/RR: 0.851/0.839, p = 0.793) and the radiological readers in the detection of basal pneumonia (cohort [3], AUC AI/reader consensus: 0.825/0.782, p = 0.390) and basal pleural effusion (cohort [3], AUC AI/reader consensus: 0.762/0.710, p = 0.336) in SCXR, partly with AUC values higher than originally published ("Nodule": 0.780, "Infiltration": 0.735, "Effusion": 0.864). The classifier "Infiltration" turned out to be very dependent on patient positioning (best in CXR, worst in SCXR). The pneumothorax SCXR cohort [2] revealed poor algorithm performance in CXRs without inserted thoracic material and in the detection of small pneumothoraces, which can be explained by a known systematic confounding error in the algorithm training process. The benefit of clinically relevant external validation is demonstrated by the differences in algorithm performance as compared to the original publication. Our multi-cohort benchmarking finally enables the consideration of confounders, different reference standards and patient positioning as well as the AI performance comparison with differentially qualified medical readers.
Subject(s)
Artificial Intelligence , Pneumothorax , Algorithms , Benchmarking , Humans , Pneumothorax/etiology , Radiography, Thoracic/methods , Retrospective StudiesABSTRACT
(1) Background: Respiratory insufficiency with acute respiratory distress syndrome (ARDS) and multi-organ dysfunction leads to high mortality in COVID-19 patients. In times of limited intensive care unit (ICU) resources, chest CTs became an important tool for the assessment of lung involvement and for patient triage despite uncertainties about the predictive diagnostic value. This study evaluated chest CT-based imaging parameters for their potential to predict in-hospital mortality compared to clinical scores. (2) Methods: 89 COVID-19 ICU ARDS patients requiring mechanical ventilation or continuous positive airway pressure mask ventilation were included in this single center retrospective study. AI-based lung injury assessment and measurements indicating pulmonary hypertension (PA-to-AA ratio) on admission CT, oxygenation indices, lung compliance and sequential organ failure assessment (SOFA) scores on ICU admission were assessed for their diagnostic performance to predict in-hospital mortality. (3) Results: CT severity scores and PA-to-AA ratios were not significantly associated with in-hospital mortality, whereas the SOFA score showed a significant association (p < 0.001). In ROC analysis, the SOFA score resulted in an area under the curve (AUC) for in-hospital mortality of 0.74 (95%-CI 0.63-0.85), whereas CT severity scores (0.53, 95%-CI 0.40-0.67) and PA-to-AA ratios (0.46, 95%-CI 0.34-0.58) did not yield sufficient AUCs. These results were consistent for the subgroup of more critically ill patients with moderate and severe ARDS on admission (oxygenation index <200, n = 53) with an AUC for SOFA score of 0.77 (95%-CI 0.64-0.89), compared to 0.55 (95%-CI 0.39-0.72) for CT severity scores and 0.51 (95%-CI 0.35-0.67) for PA-to-AA ratios. (4) Conclusions: Severe COVID-19 disease is not limited to lung (vessel) injury but leads to a multi-organ involvement. The findings of this study suggest that risk stratification should not solely be based on chest CT parameters but needs to include multi-organ failure assessment for COVID-19 ICU ARDS patients for optimized future patient management and resource allocation.
ABSTRACT
BACKGROUND: Radiology reporting of emergency whole-body computed tomography (CT) scans is time-critical and therefore involves a significant risk of pathology under-detection. We hypothesize a relevant number of initially missed secondary thoracic findings that would have been detected by an artificial intelligence (AI) software platform including several pathology-specific AI algorithms. METHODS: This retrospective proof-of-concept-study consecutively included 105 shock-room whole-body CT scans. Image data was analyzed by platform-bundled AI-algorithms, findings were reviewed by radiology experts and compared with the original radiologist's reports. We focused on secondary thoracic findings, such as cardiomegaly, coronary artery plaques, lung lesions, aortic aneurysms and vertebral fractures. RESULTS: We identified a relevant number of initially missed findings, with their quantification based on 105 analyzed CT scans as follows: up to 25 patients (23.8%) with cardiomegaly or borderline heart size, 17 patients (16.2%) with coronary plaques, 34 patients (32.4%) with aortic ectasia, 2 patients (1.9%) with lung lesions classified as "recommended to control" and 13 initially missed vertebral fractures (two with an acute traumatic origin). A high number of false positive or non-relevant AI-based findings remain problematic especially regarding lung lesions and vertebral fractures. CONCLUSIONS: We consider AI to be a promising approach to reduce the number of missed findings in clinical settings with a necessary time-critical radiological reporting. Nevertheless, algorithm improvement is necessary focusing on a reduction of "false positive" findings and on algorithm features assessing the finding relevance, e.g., fracture age or lung lesion malignancy.
ABSTRACT
(1) Background: Extracorporeal membrane oxygenation (ECMO) therapy in intensive care units (ICUs) remains the last treatment option for Coronavirus disease 2019 (COVID-19) patients with severely affected lungs but is highly resource demanding. Early risk stratification for the need of ECMO therapy upon admission to the hospital using artificial intelligence (AI)-based computed tomography (CT) assessment and clinical scores is beneficial for patient assessment and resource management; (2) Methods: Retrospective single-center study with 95 confirmed COVID-19 patients admitted to the participating ICUs. Patients requiring ECMO therapy (n = 14) during ICU stay versus patients without ECMO treatment (n = 81) were evaluated for discriminative clinical prediction parameters and AI-based CT imaging features and their diagnostic potential to predict ECMO therapy. Reported patient data include clinical scores, AI-based CT findings and patient outcomes; (3) Results: Patients subsequently allocated to ECMO therapy had significantly higher sequential organ failure (SOFA) scores (p < 0.001) and significantly lower oxygenation indices on admission (p = 0.009) than patients with standard ICU therapy. The median time from hospital admission to ECMO placement was 1.4 days (IQR 0.2-4.0). The percentage of lung involvement on AI-based CT assessment on admission to the hospital was significantly higher in ECMO patients (p < 0.001). In binary logistic regression analyses for ECMO prediction including age, sex, body mass index (BMI), SOFA score on admission, lactate on admission and percentage of lung involvement on admission CTs, only SOFA score (OR 1.32, 95% CI 1.08-1.62) and lung involvement (OR 1.06, 95% CI 1.01-1.11) were significantly associated with subsequent ECMO allocation. Receiver operating characteristic (ROC) curves showed an area under the curve (AUC) of 0.83 (95% CI 0.73-0.94) for lung involvement on admission CT and 0.82 (95% CI 0.72-0.91) for SOFA scores on ICU admission. A combined parameter of SOFA on ICU admission and lung involvement on admission CT yielded an AUC of 0.91 (0.84-0.97) with a sensitivity of 0.93 and a specificity of 0.84 for ECMO prediction; (4) Conclusions: AI-based assessment of lung involvement on CT scans on admission to the hospital and SOFA scoring, especially if combined, can be used as risk stratification tools for subsequent requirement for ECMO therapy in patients with severe COVID-19 disease to improve resource management in ICU settings.
ABSTRACT
(1) Background: Time-consuming SARS-CoV-2 RT-PCR suffers from limited sensitivity in early infection stages whereas fast available chest CT can already raise COVID-19 suspicion. Nevertheless, radiologists' performance to differentiate COVID-19, especially from influenza pneumonia, is not sufficiently characterized. (2) Methods: A total of 201 pneumonia CTs were identified and divided into subgroups based on RT-PCR: 78 COVID-19 CTs, 65 influenza CTs and 62 Non-COVID-19-Non-influenza (NCNI) CTs. Three radiology experts (blinded from RT-PCR results) raised pathogen-specific suspicion (separately for COVID-19, influenza, bacterial pneumonia and fungal pneumonia) according to the following reading scores: 0-not typical/1-possible/2-highly suspected. Diagnostic performances were calculated with RT-PCR as a reference standard. Dependencies of radiologists' pathogen suspicion scores were characterized by Pearson's Chi2 Test for Independence. (3) Results: Depending on whether the intermediate reading score 1 was considered as positive or negative, radiologists correctly classified 83-85% (vs. NCNI)/79-82% (vs. influenza) of COVID-19 cases (sensitivity up to 94%). Contrarily, radiologists correctly classified only 52-56% (vs. NCNI)/50-60% (vs. COVID-19) of influenza cases. The COVID-19 scoring was more specific than the influenza scoring compared with suspected bacterial or fungal infection. (4) Conclusions: High-accuracy COVID-19 detection by CT might expedite patient management even during the upcoming influenza season.