E³ - Young ECR Programme: Students Session
S 4 - Students Session 1
Lectures
1
Artificial intelligence in breast cancer screening: a retrospective analysis of breast cancer detection, breast cancer characteristics and workload reduction
08:00Viktor Lu, Malmö / SE
2
Impact of an artificial intelligence (AI) software on the diagnostic performance of a junior radiologist in detecting prostate cancer on MRI
08:00Ludwig Meinsohn, Rennes / FR
3
Comparison of two software versions of a novel deep-learning algorithm for haemorrhage detection in non-contrast cranial CT in a real-world patient collective
08:00Franziska Katharina Tombach, Würzburg / DE
4
A retrospective multicentre research: development and validation of deep learning classifiers for detecting inner ear malformation on temporal bone CT
08:00Abdulrahman Alkojak Almasri, Pécs / HU
5
Emergency stroke imaging: current challenges and potential solutions offered by artificial intelligence
08:00Ioana Erica Stanescu, Ploiesti / RO
6
CBCT imaging of intraosseous jaw bone lesions in children and adolescents
08:00Sara Alexia Roman, Cluj Napoca / RO
7
CT-angiography (CTA) of deep inferior epigastric artery (DIEA) and its perforators in preoperative mapping for harvesting of deep inferior epigastric perforator (DIEP) flap
08:00Evgeniia Zhigailova, Moscow / RU
8
"Radiological Cases: Quiz Time!": a new approach for the education of radiology
08:00Antonio Vitor Nascimento Martinelli Braga, Salvador / BR
9
Imaging in pulmonary embolism: a medical student's approach
08:00Violetta Benedek, Debrecen / HU
10
Volume assessment of affected lung parenchyma in COVID-19
08:00Peder Jørgensen Bruun, Bergen / NO
S 4-2
8 min
Artificial intelligence in breast cancer screening: a retrospective analysis of breast cancer detection, breast cancer characteristics and workload reduction
Viktor Lu, Malmö / Sweden
- Author Block: V. Lu, V. Dahlblom, M. Dustler, S. Zackrisson, K. Johnson; Malmö/SEPurpose: This study aimed to evaluate the diagnostic accuracy of an AI system in digital mammography (DM), analyse the detected breast cancer characteristics, and explore workload reduction.Methods or Background: Screening with DM faces challenges due to labour intensity. AI has been proposed as a solution. Double-read DM images from women who underwent screening between 2010 and 2015 in Malmö, Sweden, were analysed retrospectively and assigned an AI score between 1 and 10 by a commercial AI system (Transpara, ScreenPoint Medical). Scenario 1 assessed the diagnostic accuracy of the AI system, recalling examinations with an AI score of
- In scenario 2, women with an AI score of 10 were recalled alongside women recalled by radiologists. Proportions of various breast cancer characteristics were compared. Workload reduction was explored by excluding AI scores 1 to 9 from the radiologists' reading stream. Non-inferiority was concluded if the lower limit of the 95% CI of the difference was >5%. Results or Findings: One screening occasion per 26758 women was included, of which 753 were excluded due to failed AI analysis. Radiologists achieved a sensitivity of
- 3% (95%CI: 64.6; 76.0) and a specificity of 97.7% (95%CI: 97.6; 97.9). Scenario 1 achieved a sensitivity of 70.7% (95%CI: 65.0; 76.4) and a specificity of 91.6% (95% CI 91.3; 91.9). Scenario 2 achieved a sensitivity of 79.3% (95%CI: 74.2; 84.3) and a specificity of 90.3% (95%CI: 90.0; 90.7). The AI system and radiologists detected similar breast cancer characteristics, although the AI system missed four tubular carcinomas. The AI system could reduce the workload by 50.3% while maintaining a non-inferior sensitivity of 69.1%. Conclusion: The AI system achieved a non-inferior sensitivity but lower specificity in both scenarios, detected various breast cancer characteristics, and significantly reduced workload. Specificity could improve with consensus meetings at the expense of a minor increase in workload.Limitations: It is a retrospective study design, and prior mammograms were not analysed.Funding for this study: Funding was received from the Swedish governmental funding of clinical research (ALF).Has your study been approved by an ethics committee? YesEthics committee - additional information: This study and the data collection have been approved by the Regional Ethical Review Board in Lund University (Event No 2009/770 and 2017/326), which includes opt-out consent for the population.
S 4-3
8 min
Impact of an artificial intelligence (AI) software on the diagnostic performance of a junior radiologist in detecting prostate cancer on MRI
Ludwig Meinsohn, Rennes / France
- Author Block: L. Meinsohn1, L. Alberge2, C. Adam2, G. Herpe2, G. D'Assignies2, L. Beuzit1; 1Rennes/FR, 2Paris/FRPurpose: The purpose of this study was to evaluate the impact of an artificial intelligence (AI) software on the diagnostic performance of a junior reader (resident) in detecting clinically significant prostate cancer (csPCa) on MRI.Methods or Background: A dataset comprising 204 mpMRI cases from the PROSTATEx Challenge was employed. Each targeted lesion had a known histology. A resident read each case without AI and then after a 4-week interval with AI assistance. In contrast, an experienced radiologist reviewed the cases without AI. Their tasks encompassed lesion detection and classification per PI-RADS v
- 1 standards. The readers’ performances and the standalone AI tool were evaluated using sensitivity/specificity/accuracy metrics comparing csPCa to PI-RADS scores. PI-RADS ≥ 3 were considered as MRI positive. Differences were statistically compared using McNemmar tests. Interobserver variability was reported using Cohen’s κ on the PI-RADS scores. The reading times for all cases were also assessed and compared without and with AI. Results or Findings: The accuracy of the junior radiologist in detecting csPCa increased from
- 63 to 0.75 (p= 3.8 x10-5) with the support of AI. Sensibility and specificity also increased, respectively, from 0.84 to 0.91 (p=0.12) and from 0.53 to 0.67 (p=3x10-4). Accuracy, sensibility and specificity of the experienced radiologist were 0.78, 0.96 and 0.69, respectively, and no statistically significant difference was observed with the AI-assisted junior (p=0.5, p=0.25, p=0.86). Cohen κ between the junior and the experienced radiologist increased from 0.44 to 0.54 with the support of AI. The standalone AI accuracy, sensibility, and specificity were 0.77, 0.96, and 0.66, respectively. The annotation time of the junior radiologist was reduced by 16% using the AI tool. Conclusion: Leveraging AI notably enhanced the junior radiologist's diagnostic precision, sensitivity, and specificity in csPCa detection, mirroring the proficiency of an experienced colleague and curbing interreader discrepancies.Limitations: No limitations were identified.Funding for this study: Funding was provided by Eurostars E114567 - ProstAID.Has your study been approved by an ethics committee? YesEthics committee - additional information: The study was approved by the CHU Rennes (France) ethics committee N°
- 06.
S 4-4
8 min
Comparison of two software versions of a novel deep-learning algorithm for haemorrhage detection in non-contrast cranial CT in a real-world patient collective
Franziska Katharina Tombach, Würzburg / Germany
- Author Block: F. K. Tombach, T. A. Bley, B. M. W. Petritsch, J. F. Heidenreich, P. Gruschwitz; Würzburg/DEPurpose: The purpose of this study was to evaluate the performance of an improved deep-learning algorithm (version 2, 2022) for intracranial haemorrhage (ICH) detection in non-contrast cranial computed tomography (cCT) scans and the comparison of its performance to the previous version (version 1, 2020).Methods or Background: The deep-learning pipeline based on a three-dimensional neural network was used to automatically process cCTs. In software version 2, additional training cycles were used in an attempt to better detect subarachnoid haemorrhages in particular. A sum of n=1700 cCT, created in the period from April 2020 to April 2022, was retrospectively processed by the algorithm (version 2, 2022) and compared to the written report ("ground truth"). Discrepant results were again supervised in terms of a consensus vote by a resident with six years of experience. 519 CT scans were analysed and compared using both deep-learning software versions regarding diagnostic accuracy parameters.Results or Findings: In the clinical collective (n=1700) with a prevalence of
- 6%, the software version 2 detected ICH with a sensitivity of 94.4%, a specificity of 96.1%, an overall accuracy of 95.9% and a negative predictive value of 99.2%. In the second dataset (n=519), version 2 detected ICH with a sensitivity of 87.3%, an overall accuracy of 94.2% and a negative predictive value of 98.2%. Compared to version 1 (sensitivity=84.1%, overall accuracy=96.1%, negative predictive value=97.8%), sensitivity and negative predictive value in version 2 were increased, resulting in two fewer false negative findings. Conclusion: The improvements in the deep-learning software version 2 lead to a significant increase in sensitivity and negative predictive value and, thus, fewer false negative findings when used as a triage tool.Limitations: Limitations included the retrospective study design and the single-centre, single-vendor approach. We clinically indicated cCT without preselection. A supervised (resident with six years experience) radiology report was used as ground truth.Funding for this study: Philipp Gruschwitz [Grant No Z-02CSP/18] was funded by the Interdisciplinary Center of Clinical Research Würzburg, Germany. The Department of Diagnostic and Interventional Radiology received a Siemens research grant. The other authors of this manuscript declare no relationships with any companies, whose products or services may be related to the subject matter of the article.Has your study been approved by an ethics committee? YesEthics committee - additional information: The study received permission from the local institutional review board (Ethic Committee of the University of Würzburg; protocol number: 20230919 02).
S 4-5
8 min
A retrospective multicentre research: development and validation of deep learning classifiers for detecting inner ear malformation on temporal bone CT
Abdulrahman Alkojak Almasri, Innsbruck / Austria
- Author Block: A. A. Almasri1, A. Dhanasingh2, S. Sugarova3, F. Almuhawas4, L. Hofmeyr5, F. Wagner6, E. Wappl-Kornherr2, E. Kedves7, A. Alsanosi4, K. Sriperumbudur2, A. Kedves2; 1Pécs/HU, 2Innsbruck/AT, 3St. Petersburg/RU, 4Riyadh/SA, 5Stellenbosch/ZA, 6Bern/CH, 7Sopron/HUPurpose: The ability to identify inner ear malformations (IEM) has been demonstrated by deep learning (DL) and artificial intelligence (AI). Based on patients' computed tomography (CT), we created an automated system to identify a specific IEM.Methods or Background: While developing the deep learning model for inner ear CTs used in this retrospective, multicentre study, we included 2053 patients who had been imaged between 2016 and
- Three nations - Saudi Arabia, South Africa, and Russia - provided temporal CT datasets. Deep convolutional neural networks were used to create supervised learning models, and all of the data were categorised as incomplete partitiontype III or other. 25 professional experts with or without training from Austria, the United Kingdom, South Africa, and Egypt evaluated 24 patients for interobserver validity by covering the variability of observers.Results or Findings: The specificity and sensitivity of supervised learning models were
- 1%, 88.4%, 80.6%, and 88.1%, respectively. The performance of the two-stage DL algorithm was better than the one-stage algorithm (AUC
- 86, 95% CI 0.82-0.90; AUC 0.80, 95% CI 0.74-0.86). Interobserver analysis using Kruskal Wallis ANOVA and one sample Wilcoxon test revealed that the profession (including AI) had an impact on correctly identifying present or absent malformations but not training (p=0.0674). The analyses even showed that the correct assignment by AI was superior to professionals (p=0.0403). Conclusion: We outline the development and verification of a potential fully automated workflow for IEM detection. The decision-maker must supervise the tool, even though it may have good diagnostic accuracy when risk stratification is being done.Limitations: The limitation of this study is the possible imbalance, which may occur and could cause overfitting.Funding for this study: No funding was provided for this study.Has your study been approved by an ethics committee? YesEthics committee - additional information: This study was approved by the independent ethics committee of 3 Hospitals IRB Nos. 22/0084/IRB, 23_001/IRB, and S_23_001/IRB, respectively.
S 4-6
8 min
Emergency stroke imaging: current challenges and potential solutions offered by artificial intelligence
Ioana Erica Stanescu, Frankfurt am Main / Germany
Author Block: L. Plamadeala, I. E. Stanescu; Iasi/RO
Purpose: The study aims to provide a comprehensive overview of recent advancements in artificial intelligence (AI), machine learning (ML), and advanced imaging techniques in the context of stroke care. It synthesizes findings from the most recent 55 studies, sourced from PubMed and conducted between 2016 and 2023, to elucidate the transformative potential of these technologies across various facets of stroke management.
Methods or Background: A systematic review of these studies explores AI and ML applications in stroke care, including diagnostic accuracy, treatment optimisation, imaging enhancement, prognosis prediction and lesion segmentation. Diverse methodologies, such as convolutional neural networks, support vector machines, deep learning models and motion correction algorithms, are employed. Data from these studies are analysed to assess the impact and effectiveness of these technologies in stroke management.
Results or Findings: The findings collectively reveal the profound impact of AI and ML technologies on stroke care. They enable rapid and precise diagnosis, efficient treatment selection, enhanced imaging interpretation, accurate prognosis prediction, and sensitive lesion segmentation. Convolutional neural networks and support vector machines exhibit remarkable efficiency in stroke subtype identification and large vessel occlusion detection. Furthermore, motion correction algorithms improve image quality and lesion detectability in cerebral CT, while deep learning models predict stroke using raw EEG data with exceptional accuracy. Automation platforms for intracranial large vessel occlusion detection expedite diagnostic work-ups, and multimodal deep learning frameworks like DeepStroke outperform traditional triage methods.
Conclusion: The fusion of AI, ML, and advanced imaging transforms stroke care and enhances diagnosis, treatment, imaging, and prognosis. To fully benefit, we must tackle research gaps in treatment studies and address data privacy and integration challenges.
Limitations: When implementing these innovative stroke care approaches, we must address some key limitations, including a focus on diagnosis, modality-specific strategies, and data-related challenges.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study is educational.
Purpose: The study aims to provide a comprehensive overview of recent advancements in artificial intelligence (AI), machine learning (ML), and advanced imaging techniques in the context of stroke care. It synthesizes findings from the most recent 55 studies, sourced from PubMed and conducted between 2016 and 2023, to elucidate the transformative potential of these technologies across various facets of stroke management.
Methods or Background: A systematic review of these studies explores AI and ML applications in stroke care, including diagnostic accuracy, treatment optimisation, imaging enhancement, prognosis prediction and lesion segmentation. Diverse methodologies, such as convolutional neural networks, support vector machines, deep learning models and motion correction algorithms, are employed. Data from these studies are analysed to assess the impact and effectiveness of these technologies in stroke management.
Results or Findings: The findings collectively reveal the profound impact of AI and ML technologies on stroke care. They enable rapid and precise diagnosis, efficient treatment selection, enhanced imaging interpretation, accurate prognosis prediction, and sensitive lesion segmentation. Convolutional neural networks and support vector machines exhibit remarkable efficiency in stroke subtype identification and large vessel occlusion detection. Furthermore, motion correction algorithms improve image quality and lesion detectability in cerebral CT, while deep learning models predict stroke using raw EEG data with exceptional accuracy. Automation platforms for intracranial large vessel occlusion detection expedite diagnostic work-ups, and multimodal deep learning frameworks like DeepStroke outperform traditional triage methods.
Conclusion: The fusion of AI, ML, and advanced imaging transforms stroke care and enhances diagnosis, treatment, imaging, and prognosis. To fully benefit, we must tackle research gaps in treatment studies and address data privacy and integration challenges.
Limitations: When implementing these innovative stroke care approaches, we must address some key limitations, including a focus on diagnosis, modality-specific strategies, and data-related challenges.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study is educational.
S 4-7
8 min
CBCT imaging of intraosseous jaw bone lesions in children and adolescents
Sara Alexia Roman, Cluj Napoca / Romania
- Author Block: S. A. Roman, E. Crasnean, A. Ban, M. Hedesiu, R. A. Roman; Cluj Napoca/ROPurpose: To present in a Cone Beam Computer Tomography (CBCT) pictorial the imaging features of the jaw bone space-occupying lesions in children and adolescents, both benign and malignant ones.Methods or Background: Intraosseous lesions of the jaw in children are not very common, mosly being asymptomatic. Several of the benign ones may present locally aggressive features and need proper recognition and management. Lesions are found often incidentally on panoramic radiography or CBCT, indicated by dentists or orthodontists for abnormalities in teeth eruption. We present retrospective CBCT examinations of under 18 years old patients treated for jaw bone masses in the last 5 years in the Maxillofacial Surgery Clinic, ages ranging from 5 to
- By using individualized CBCT reconstructions, imaging characteristics were analyzed and correlated with histopathology. Several features were assessed: density, internal structure, contour, locularity, location, the relationship with nearby teeth and structures, effect on corticals, e.g. Results or Findings: Differences and similarities between imaging characteristics are presented, structured for odontogenic and non-odontogenic, benign and malignant. Images for different cystic lesions, being the majority of the cases treated, are presented, from cysts like naso-palatal, inflammatory radicular, dentigerous, the typical and atypical odontogenic keratocyst, to central giant cell granuloma, cherubism or ameloblastoma, followed by the denser lesions, such as cementoblastoma, cementifying or ossifying fibromas and odontomas. Malignancy was present in the group, mostly represented by osteosarcoma.Conclusion: Most of the jawbone lesions in children are benign entities, with typical CBCT features, the mandible being the most affected site. Atypical presentation may pose problems in reaching a diagnosis, and needs attention, since malignant lesions are not uncommon. The option of individualized sections in CBCT helps narrowing the differential diagnosis, allowing a proper morphological lesion evaluation.Limitations: No limitations were identified.Funding for this study: No funding was received for this study.Has your study been approved by an ethics committee? Not applicableEthics committee - additional information: Not applicable
S 4-8
8 min
CT-angiography (CTA) of deep inferior epigastric artery (DIEA) and its perforators in preoperative mapping for harvesting of deep inferior epigastric perforator (DIEP) flap
Evgeniia Zhigailova, Moscow / Russia
- Author Block: Z. Evgeniia, E. Mershina, V. Sinitsyn; Moscow/RUPurpose: The aim of the study was to define the capabilities of CTА in the preoperative mapping of DIEA and its perforators for harvesting DIEP flap.Methods or Background: DIEP flaps have become the method of choice for autologous breast reconstruction. The study of variable anatomy and preoperative mapping of DIEA and its perforators using CTA are vital for optimal flap harvesting. We retrospectively analysed CTA of the abdominal aorta and its branches of 20 patients (m/f - 7/13, mean age 53±12 yrs). CTA was performed with dual-source CT. We used the Mann-Whitney test for continuous outcomes and Spearman’s correlation for binary outcomes with p≤
- 05 for significance. 3D models with cinematic-rendering techniques were created. Results or Findings: CTA datasets from 20 patients with 104 perforators were analysed. The DIEA diameters varied from
- 76 to 1.93 mm (mean 1.27±0.24 mm), the distance from the vessel to the linea alba varied from 1.25 to 8.74 cm (4.46±1.52 cm), and the distance to the umbilicus varied from 2.81 to 9.73 cm (5.58±1.37 cm). A statistically significant correlation was found between the diameter of DIEA and the quantity of perforators (p=0.575). There was no statistically significant correlation between DIEA’s diameter and its perforators’ diameters (p=0.233). No correlations were found between DIEA’s diameter and the diameter of iliac arteries (p1=0.091; p2 =-0.051; p3=0.049) and between the presence of calcinosis in the common iliac artery (U=43, p>0.05). Conclusion: This is one of a few studies to analyse the use of preoperative mapping of DIEA for harvesting DIEP flaps with CTA. The results demonstrated the usefulness of CTA and 3D reconstruction in studying DIEA and DIEP variable anatomy for preoperative mapping.Limitations: The study's limitations are its retrospective nature and the insufficient number of patients.Funding for this study: No funding was received for this study.Has your study been approved by an ethics committee? Not applicableEthics committee - additional information: No information was provided by the submitter.
S 4-9
8 min
"Radiological Cases: Quiz Time!": a new approach for the education of radiology
Antonio Vitor Nascimento Martinelli Braga, Salvador / Brazil
- Author Block: A. V. N. Martinelli Braga, A. L. Costa, C. F. Lins; Salvador/BRPurpose: In this study, we aim to analyse medical students' perception regarding the "Radiological Cases: Quiz Time!" gamification tool as a novel approach to radiology education.Methods or Background: The gamification tool "Radiological Cases: Quiz Time!" was elaborated by radiology monitors guided by a radiologist using the "Kahoot" online platform. The class was divided into groups, and multiple questions regarding radiological cases were asked. Then, students were invited to answer a questionnaire that assessed their sociodemographic profile, self-assessment of learning radiological anatomy and the student's opinion about the radiological workshop "Radiological Cases: Quiz Time!". The questions were based on the Likert-modified scale, totalling 35 questions. Incomplete questionnaires were excluded. The Cronbach's alpha analysis was performed, and values above 0,7 were considered acceptable.Results or Findings: Of the 150 invited, 59 completed the questionnaire, averaging 20±5 years old and from those, 34 (57,62%) were women. Cronbach's alpha was 0,94, attesting to the internal consistency of the study. In total, 44 (74,57%) considered that the gamification tool contributed substantially to learning; 48 (81,35%) considered the workshop to have high educational value; 44 (74,57%) judged the material as clinically significant for their future clinical practice/experience; and 53 (89,83%) considered that the "Radiological Cases: Quiz Time!" contributed to form friendship bonds.Conclusion: The workshop "Radiological Cases: Quiz Time!" is an interactive, creative and innovative tool for teaching radiology, promoting clinically meaningful learning and the formation of new bonds of friendship.Limitations: The study's main limitation was the low response to the questionnaire as it was administered at the end of the semester, close to the holidays.Funding for this study: No funding was received for this study.Has your study been approved by an ethics committee? YesEthics committee - additional information: The study received institutional review board approval, and written informed consent was obtained from all participants. CAAE:
- 2.0000.5544.
S 4-10
8 min
Imaging in pulmonary embolism: a medical student's approach
Violetta Benedek, Debrecen / Hungary
- Author Block: V. Benedek; Debrecen/HUPurpose: Pulmonary embolism is a potentially fatal status, and diagnosing and treatment needs an institutional environment. For a medical student, it is important to be familiar with the most common pathologies and available diagnostic and therapeutic techniques. We should also be aware of limitations and potential harms. Proper referrals and awareness of the ALARA principle are essential for every physician. My goal was to evaluate chest CT referrals in pulmonary embolism.Methods or Background: Chest CT angiography studies were selected during August and September 2022 in two emergency centres at the University of Debrecen, Hungary. CT scans, D-dimer levels and other clinical findings (including pregnancy, known malignancy or inflammatory disease) were collected in a database. The data of 380 patients (171 male and 209 female) was evaluated. Additionally, 905 chest CT studies using different imaging protocols were evaluated to compare the applied doses in the same period.Results or Findings: Elevated D-dimer (>0,5 μg/ml) values were measured in 344 (90,52%) cases, and 32 patients (9,3%) were diagnosed with pulmonary embolism. Only three patients had pulmonary embolism without elevated D-dimer value. A comparison of applied doses among different chest CT protocols showed that pulmonary embolism protocol study doses are significantly lower than in other protocols.Conclusion: Pulmonary embolism is a commonly referred question by emergency departments, while positivity or the ratio of other pathology findings is low during imaging. When requesting these examinations, it is crucial to consider the patient's clinical status, various scoring systems, and additional diagnostic tests. The role of radiologists and radiographers is to design appropriate imaging protocols and workflows supporting referring physicians in finding the most appropriate diagnostic tool for their patient's clinical condition. The role of proper referrals should be emphasised in undergraduate training.Limitations: Only limited clinical conditions were considered during the patient status evaluation.Funding for this study: No funding was received for this study.Has your study been approved by an ethics committee? YesEthics committee - additional information: The study was approved with the registration number: DE KK RKEB.IKEB. 6291-
S 4-11
8 min
Volume assessment of affected lung parenchyma in COVID-19
Peder Jørgensen Bruun, Bergen / Norway
- Author Block: P. J. Bruun, A. P. Parkar; Bergen/NOPurpose: This study aims to investigate if quantitative CT volume assessment as a stand-alone criterion of affected lung parenchyma is helpful in distinguishing non-critically ill and critically ill patients with COVID-Methods or Background: Patients admitted between March 2020 and December 2021 with an RT-PCR-confirmed COVID-19 diagnosis and a chest CT were included retrospectively. Patients were divided into two groups: critically ill (patients who died or were admitted to the ICU) and non-critically ill for the others. The percentage of affected lung parenchyma was analysed by two observers using a semi-quantitative method. The volumes and the basic demographic data were collected, and reliability between the observers was assessed. Statistical analyses were done in SPSS.Results or Findings: 67 patients (41 males and 26 females) were included. Eleven patients were admitted to the ICU, and five died without being admitted to the ICU. The mean volume of affected lungs was 35% in males and 31% in females. Lung volume affection over 60% led to ICU admittance, but lower values did not exclude ICU admittance. Depending on which cut-off value was used to distinguish the groups (20%-40%-60%), the sensitivity was 88%-79%-80%, but the specificity was low at 26%-24%-32%. The PPV increased 24%-75%-100%, but the NPV dropped 88%-38%-19%. A cut-off to distinguish the two groups could not be determined. Reliability between the two observers was good (ICC=
- 8). Conclusion: Semi-quantitative CT-volume assessment of affected lung parenchyma as a stand-alone criterion is not helpful in distinguishing between non-critically ill and critically ill patients with COVID-
- The use of lung CT for evaluating the severity of the disease is not advisable. Limitations: Analysers were not completely blinded to all patient outcomes. The time between admittance and CT was not standardised.Funding for this study: No funding was received for this study.Has your study been approved by an ethics committee? YesEthics committee - additional information: The study was approved by the Norwegian regional ethics committee (REK).