Research Presentation Session: Chest

RPS 1404 - From diagnosis to prognosis: chest imaging

March 1, 12:30 - 13:30 CET

7 min
AI vs senior radiologists in detecting thoracic abnormalities on chest radiographs compared to CT
Nor-Eddine Regnard, Paris / France
Author Block: S. Bennani1, J. Ventre1, V. Marty1, E. Lacave1, D. Hayashi2, A. Kompel3, A. Gupta3, A. Guermazi4, N-E. Regnard1 ; 1Paris/FR, 2Stony Brook, NY/US, 3Boston, MA/US, 4West Roxbury, MA/US
Purpose: The study aimed to assess the diagnostic performances of an artificial intelligence (AI) software in the detection of thoracic abnormalities on chest radiographs compared to senior radiologists.
Methods or Background: We collected 319 chest radiographs of patients above 22 years old who underwent thoracic CT within 72 hours. A senior chest radiologist annotated the radiographs for four abnormality types (pleural abnormality, consolidation, mediastinal-hilar abnormality, nodule) using CT findings as the ground truth. Three senior radiologists independently analysed the dataset, knowing clinical indications without CT access. Discrepancies were resolved by consensus.
The AI (ChestView, Gleamer), a deep learning algorithm that detects the four abnormalities, was compared against the radiologists and their consensual analysis for sensitivity and specificity.
Results or Findings: The dataset included 168 radiographs (age: 64±16 years, 91 women): 129 with at least one abnormality, 39 without any abnormality.
For consolidation, the sensitivities were 72% for AI, 54%, 80%, and 66% for the individual readers, and 71% for consensus, with specificities of 92% for AI, 80%, 77%, 85% for the readers, and 92% consensus. The sensitivities for mediastinal-hilar abnormalities, were 54% (AI), 43%, 27%, 48% (readers), and 54% (consensus); specificities were 95%, 88%, 96%, 93%, and 94%, respectively. For nodules, the sensitivities were reported as 67% for AI, 57%, 55%, 55% for the readers, and 62% consensus, with specificities of 89%, 52%, 88%, 81%, and 86%, respectively. Lastly, for pleural abnormalities, the sensitivities were 84%, 89%, 87%, 73%, and 83%, and the specificities were 93% for AI, 88%, 92%, 95% for the readers, and 94% for consensus.
Conclusion: The AI consistently matched or exceeded senior radiologists in detecting thoracic abnormalities.
Limitations: The dataset had few radiographs with high pathology prevalence.
Funding for this study: Funding for this study was received from Gleamer.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by WCG (number IRB00000533).
7 min
Impact of the recent global iodinated contrast agent shortage on positivity rate for pulmonary embolism in CT pulmonary angiograms and chest CT with contrast at two major US health care systems
Axel Wismueller, Pittsford / United States
Author Block: A. Wismueller1, J. P. Kanne2, L. Stockmaster1, E. Weinberg1, D. Shrier1, A. Vosoughi1, A. Kasturi1, N. Hadjiyski1; 1Rochester, NY/US, 2Madison, WI/US
Purpose: The objective of this study was to quantitatively track the impact of the recent global shortage of iodinated contrast agents on pulmonary embolism (PE) positivity rate in CT pulmonary angiograms (CTPA) using AI-based image analysis at two major US healthcare systems.
Methods or Background: Using commercial AI-based image analysis (Aidoc Medical), we analysed daily volumes, PE and incidental PE (iPE) positivity rates for 7,633 computed tomography pulmonary angiogram (CTPA) and 11,164 contrast-enhanced chest CT (CT+C) exams before and during the contrast agent shortage (both comprising 01.04.2022 through 01.07.2022). For comparison, we analysed daily volumes and positivity rates for intracranial haemorrhage (ICH) on 30,803 non-contrast head CT exams during the same period. We compared two observational periods: a pre-shortage control period (10.04.2022 through 40.04.2022), and a contrast shortage period (20.05.2022 through 10.06.2022). Percentage change metrics of case volumes and positivity rates for PE, iPE and ICH were calculated.
Results or Findings: Case volumes of CTPA exams dropped from baseline during the shortage period by 38.9% while PE positivity rates significantly increased by 31.1% (p< 10^-4, Welch's unequal variances t-test). Similarly, case volume of CT+C exams dropped during the contrast agent shortage period by 49.9% while iPE positivity rates significantly increased by 20.2%, (p< 10^-4). For comparison, non-contrast head CT volumes dropped by only 6.5%, and ICH positivity rates increased by only 15.9%, with no significant difference (p> 0.05).
Conclusion: Our results suggest a significant increase of PE/iPE positivity rates in significantly decreased CTPA/CT+C total exam volumes during the observed global contrast agent shortage period, while non-contrast head CT exam volumes and ICH positivity rates remained essentially stable. Our observations can be explained by more restrictive ordering patterns for contrast-enhanced studies during the shortage period.
Limitations: Ground truth was based on AI image analysis rather than NLP on radiology reports.
Funding for this study: Funding was received from the ACR Innovation Award for AW; Aidoc Medical.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by URMC and UWHealth institutional IRB approvals.
7 min
Correlation between postoperative pulmonary venous congestion on chest radiograph and outcome after cardiac transplantation
Charlotte Sonja Böttger, Düsseldorf / Germany
Author Block: C. S. Böttger, C. B. Fink, V. Hettlich, F. S. Jenkins, D. Scheiber, H. Aubin, A. Lichtenberg, S. Reinartz, U. Boeken; Düsseldorf/DE
Purpose: After heart transplantation (HTX), chest radiography plays an important role in the simple, non-invasive detection of pathological changes. Given the significant relevance of fluid balance following this procedure, particularly for hemodynamic regulation, the assessment of potential pulmonary venous congestion (p.v.-congestion) in chest radiographs is crucial. The objective of this analysis was to determine a potential correlation between early postoperative p.v.- congestion after cardiac transplantation and subsequent outcomes.
Methods or Background: 302 patients underwent HTX between 09.2010 and 09.2023 in our department. Recipients were retrospectively divided into 3 groups based on the severity of signs of pulmonary venous congestion on chest x-ray on the first postoperative day: Group 1 had no signs of p.v.-congestion (n= 115), Group 2 showed signs of mild p.v.- congestion (n= 145) and Group 3 showed signs of moderate to higher-grade p.v.- congestion (n= 41). One patient was excluded due to intraoperative death.
Results or Findings: The 30-day mortality was 4.5% in group 1, 6.2% in group 2, and 22% in group 3 (p< 0.05). Combined donor heart/lung organ harvesting was performed significantly more often in group 2 compared to group 3. Intraoperative bypass time and duration of operation were shorter in groups 1 and 2 compared to group 3. There were no significant differences between groups regarding graft rejection or infection. Patients in group 3 developed significantly more neurological complications postoperatively. 1-year and 3-year survival was reduced in patients with signs of moderate to higher grade congestion on X-ray (p> 0.05).
Conclusion: In this analysis, it was demonstrated that the presence of signs of p.v.-congestion and the severity on the first postoperative day after HTX have an influence on survival. Therefore, a conservative estimate of survival can be made by assessing the first postoperative chest radiograph.
Limitations: This was a single centre study.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Not applicable.
7 min
Comparison of two CT-derived methods to assess alveolar collapse as a potential prognostic imaging marker in pulmonary fibrosis
Hoen-oh Shin, Hannover / Germany
Author Block: S. Scharm1, C. M. Schaefer-Prokop2, A. Schreuder3, J. Fuge1, F. Wacker1, B. Seeliger1, A. Prasse1, H-o. Shin1; 1Hannover/DE, 2Amersfoort/NL, 3Nijmegen/NL
Purpose: The objective of this study was to evaluate two different CT-derived approaches for the assessment of alveolar collapse, a potential precursor of pulmonary fibrosis.
Methods or Background: For this single-centre retrospective longitudinal study, 66 consecutive patients with idiopathic pulmonary fibrosis underwent CT in inspiration and expiration and pulmonary function testing at baseline. The patient population was divided into two subgroups according to their status at 3 years (death or transplantation versus clinical surveillance). Parametric response maps were generated as scatterplots of voxel-wise attenuation values of paired inspiration and expiration scans after non-linear registration. Inspiratory and expiratory attenuation histograms were also generated and analysed. Voxels with an abnormally high increase in attenuation during expiration were interpreted as "collapsed" lung tissue.
A Mann-Whitney U test was performed to assess the difference in CT-derived measures between the two subgroups and logistic regression was performed to test the predictive power of FVC and both CT-derived measures (PRM and histogram analyses).
Results or Findings: All CT-derived PRM and histogram measures were significantly different (p < 0.005) between the two patient subgroups.
With a discriminatory performance of AUC = 0.788, 95% CI 0.679-0.898 for PRM and AUC = 0.791, 95% CI = 0.684- 0.899 these two CT-derived measures had a superior predictive performance compared to FVC alone with AUC = 0.708, CI 95% 0.581- 0.836.
The advantage of the PRM analysis is a voxel-wise matching between inspiration and expiration due to the required registration. On the other hand, the histogram parameter approach does not require advanced postprocessing but lacks the direct spatial correlation of the attenuation change between inspiration and expiration.
Conclusion: Both methods perform similarly well in predicting patient outcome by directly or indirectly quantifying alveolar collapse, and both outperform FVC.
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? Yes
Ethics committee - additional information: This study was approved by IRB, with the approval code: No. 10726_BO_K_2023
7 min
Effect of inspiratory lung volume on bronchial and arterial dimensions on chest CT
Yuxin Chen, Rotterdam / Netherlands
Author Block: Y. Chen1, R. Latisenko2, D. Lynch3, P. Ciet1, J-P. Charbonnier2, H. A. W. M. Tiddens2; 1Rotterdam/NL, 2Nijmegen/NL, 3Denver, CO/US
Purpose: Bronchus-artery (BA) dimensions on chest CT are influenced by inspiratory lung volume. We aimed to assess the effect of inspiratory lung volume on BA-ratios in patients with COPD.
Methods or Background: A selected group participating in the COPDGene study received a full dose (FD, 120kVp, 200mAs) and a reduced dose (RD, 120kVp 35mAs) CTs in the same imaging session. CTs were analysed using LungQ. For segmental (G0) and distal generations, the following diameters were measured: bronchial outer (Bout), inner (Bin), wall thickness (Bwt), and artery (A), and the following BA-ratios were computed: Bout/A, Blumen/A, Bwt/A, and bronchial wall area/bronchial outer area (Bwa/Boa). Total lung capacity of the CT (TLC-CT) was computed. Differences between the volumes (ΔTLC-CT%) between the two CTs were expressed as % of the highest TLC-CT. Mixed-effect models were used to investigate the influence of TLC-CT on BA-ratios adjusted for dose protocol. AUC was used to define a cut-off value for BA-ratios in relation to ΔTLC-CT%.
Results or Findings: 1319 patients with a mean (SD) age of 64.4 (8.7) years were included. 329 (124) BA-pairs could be analyzed per CT. No significant difference was found for TLC-CT between FD and RD scans. A ΔTLC-CT% of 5% significantly lead to 0.015 and 0.025 decreases in Bout/A and Bin/A and 0.004 and 0.015 decreases in Bwt/A and Bwa/Boa. The cut-off values to determine when BA-ratios were influenced by lung volume were for Bout/A 4.9%, Bin/A 3.7%, Bwt/A 4.4%, and Bwa/Boa 5.4%.
Conclusion: The BA biomarkers provide a robust quantification of bronchial changes on CT when differences in ΔTLC-CT% are kept below 5%. Standardising volumes for clinical follow-up and trials is recommended to optimise assessment for tracking airway disease changes over time.
Limitations: The study examined a selected group of an elderly population,
Funding for this study: This study was funded by the Erasmus MC LungAnalysis supporting grant.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The COPDGene study was approved by the institutional review boards at each of the 21 participating clinical sites.
7 min
Quantitative CT characterises baseline regional structure and function in idiopathic pulmonary fibrosis patients with one-year diffusing capacity decline
Hongseok Ko, Seoul / Korea, Republic of
Author Block: H. Ko1, W. C. Chung2, J. Choi3, S. M. Choi2, C-H. Lee2, K. J. Chae4, C. H. Lee2; 1Chuncheon/KR, 2Seoul/KR, 3Kansas City, KS/US, 4Jeonju/KR
Purpose: We characterised regional lung structure and function in idiopathic pulmonary fibrosis (IPF) patients with a one-year diffusing capacity of carbon monoxide (DLCO) decline using quantitative inspiratory and expiratory chest computed tomography (CT) analysis.
Methods or Background: Baseline and one-year pulmonary function tests (PFTs) and baseline inspiration-expiration CT scans were collected from 51 IPF patients (71.5±5.9 years). Commercial and in-house quantitative CT (qCT) software were used for the segmentation and quantification of 113 regional lung structural-functional features. Baseline qCT features were compared between patients with 5% or more percent-predicted (%pred) DLCO decline (IPF-A) and the rest (IPF-B), using Wilcoxon rank-sum test.
Results or Findings: Compared to IPF-B (n= 25, 1 female), IPF-A (n= 26, 4 females) had less baseline high attenuation area percent (HAA%) in the right upper (RUL, -32.0%, p= 0.011) and lower (RLL, -29.3%, p= 0.027) lobes and greater anisotropic deformation index (31.2%, p= 0.030) and relative displacement (26.2%, p= 0.047) in the left lower lobe (LLL). IPF-A also had moderately greater normalised hydraulic diameter (Dh*) at the RLL segmental airways (sRLL) (9.3%, p= 0.054). Early traction bronchiectasis was observed in the sRLL of IPF-A. IPF-A showed CT features of greater extent and severity of fibrosis in IPF. Demographics and PFTs were not significantly different at baseline.
Conclusion: qCT analysis characterised baseline regional lung structure and function in IPF with DLCO decline and supported visual CT interpretation. Findings suggest that advanced fibrotic changes and RLL segmental airway traction bronchiectasis may be precursors of DLCO decline presumably with progressive pulmonary fibrosis.
Limitations: This study has a limited number of cases.
Funding for this study: This study was supported by Korea Environmental Industry & Technology Institute (KEITI) grant 2018001360001 funded by the Ministry of Environment (MOE), Republic of Korea.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study protocol was approved by the institutional review boards (IRB) of Seoul National University Hospital (IRB: 1810-036-977)
7 min
Default vs selective application of artificial intelligence to classify viral pneumonia on CT imaging: impact on human performance
Francesco Rizzetto, Rho / Italy
Author Block: F. Rizzetto, L. Berta, L. A. Carbonaro, D. Artioli, F. Travaglini, G. Zorzi, A. Torresin, P. E. Colombo, A. Vanzulli; Milan/IT
Purpose: This study aimed to assess the impact of different artificial intelligence (AI)-powered approaches on human performance in distinguishing COVID-19 from other viral pneumonia on CT imaging.
Methods or Background: Three experienced radiologists blindly evaluated 220 chest CT examinations of patients with viral pneumonia (n=151 COVID-19; n=69 other viruses), assigning a CO-RADS score before (S1) and after (S2) receiving results from a validated AI classifier. Inter-reader agreement with Gwet’s agreement coefficient type-2 (AC2) and performance metrics were calculated for S1 and S2. Two different S2 scenarios were considered: one where the AI output was available for all cases (default approach), and another where the AI prediction was applied only to CO-RADS=3 cases from S1 (selective approach).
Results or Findings: The readers showed good-to-excellent agreement in assigning CO-RADS for all scenarios (range AC2=0.77-0.81). On average, CO-RADS changes between S1 and S2 occurred in 18% of cases, with 29% involving patients initially assigned CO-RADS=3. In these cases, the use of AI output correctly classified 85% of patients. Conversely, when the radiologists were confident in S1 diagnosis (CO-RADS≠3), class changes in S2 occurred in 7% of cases. This prevented incorrect diagnosis in 45% of patients but led to a missed correct classification in the remaining 55%. In S1, the readers achieved 78% accuracy, with 15% of patients classified as CO-RADS=3. In S2, accuracy was 81% with 16% of CO-RADS=3 for the default approach, and 79% with 10% of CO-RADS=3 for the selective approach. Significant differences were only observed in the proportion of CO-RADS=3 cases in S2 with the selective approach (p<0.009).
Conclusion: AI helps distinguish COVID-19 from other viral pneumonias in equivocal cases but its reliability diminishes when the reader is already confident on diagnosis.
Limitations: The main limitation is the retrospective single centre study design.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study received approval from the Local Ethics Committee (decision number: 188-22042020)

This session will not be streamed, nor will it be available on-demand!