Research Presentation Session: Artificial Intelligence & Machine Learning & Imaging Informatics

RPS 2105 - Exploring the latest frontiers in machine learning techniques for MDCT advancements

March 2, 16:00 - 17:30 CET

6 min
Moderator introduction
Rebeca Mirón Mombiela, Herlev / Denmark
7 min
Coronary artery calcium scoring on the segment-level using deep active multi-task learning for time-efficient annotation
Bernhard Föllmer, Berlin / Germany
Author Block: B. Föllmer1, S. Tsogias1, F. Biavati1, M. Bosserdt1, K. F. Kofoed2, P. Maurovich-Horvat3, P. Donnelly4, T. Benedek5, M. Dewey1; 1Berlin/DE, 2Copenhagen/DK, 3Budapest/HU, 4Belfast/UK, 5Targu Mures/RO
Purpose: The aim of this study was to develop and evaluate a time-efficient annotation strategy for multi-task segment-level coronary artery calcium scoring (CACS) on non-contrast CT for the improvement of localisation and quantification of calcifications in the coronary artery tree.
Methods or Background: This study included 1514 patients (mean age 60.0 ± 10.2 years, 55.7% female) with stable chest pain from the multicentre DISCHARGE trial (NCT02400229), which were randomly divided into a training/validation set (1514), and a test set (455). We developed a deep active learning strategy for time-efficient annotation of coronary artery segment-regions for auxiliary task learning in a multi-task model for segmentation of CACs on segment-level. We compared the model with a baseline U-Net in terms of micro-average sensitivity and micro-average specificity for assigning detected calcification to the correct segment and analysed interobserver variability in a subset of 150 patients.
Results or Findings: The micro-average sensitivity and micro-average specificity for assigning detected calcification to the correct segment improved from 0.581 (95% CI, 0.550-0.613) to 0.732 (95% CI, 0.711-0.754, p<0.001) and from 0.965 (95% CI, 0.962-0.968) to 0.978 (95% CI, 0.976-0.980), respectively (p<.001), compared to the baseline model, with an additional annotation time of less than 12 hours for annotation of coronary artery segment-regions. The agreement between the model and the reference standard (first observer) for segment class assignment was good with a weighted Cohen’s κ of 0.806 (95% CI, 0.782-0.828) and only slightly lower compared to the second observer (weighted Cohen’s κ of 0.819 [95% CI, 0.789-0.849] (p<.001).
Conclusion: Deep active learning can be used for time-efficient annotation of coronary artery regions to improve the performance of a multi-task model for segment-level CAC scoring.
Limitations: A consensus between the two observers that can serve as a reference standard was not available.
Funding for this study: This work was funded by the German Research Foundation through the graduate program BIOQIC (GRK2260, project-ID: 289347353), and the DISCHARGE project (603266-2, HEALTH-2012.2.4.-2) funded by the FP7 Program of the European Commission.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The DISCHARGE trial was approved by the ethics committee at Charité–Universitätsmedizin Berlin as the coordinating center, by the German Federal Office for Radiation Protection, and by local or national ethics committees.
7 min
Auto-LSN: fully automated liver surface nodularity quantification for the diagnosis of advanced fibrosis in CT imaging
Sisi Yang, Paris / France
Author Block: S. Yang1, A. Bône2, T. Decaens3, A. J. Glaunes1; 1Paris/FR, 2Villepinte/FR, 3Grenoble/FR
Purpose: The liver surface nodularity score was proposed in 2016 to measure the irregularities of the surface of the liver in imaging and was shown to positively correlate with fibrosis, graded by the Metavir score. The LSN software requires the user to manually draw regions of interest before automatically segmenting the liver contour and measuring the score. We propose an entirely automated alternative (auto-LSN), and we compare its potential for the diagnosis of advanced fibrosis with LSN.
Methods or Background: It was a monocentric retrospective study on portal phase CT images in patients with suspected primary hepatic tumors and an available Metavir score between 2015 and 2020.
With respect to LSN, our auto-LSN method innovates on three key components: (i) automatic segmentation of the liver contour with deep learning, (ii) exclusion of low-contrast, high-curvature or inner regions of this liver contour, (iii) refinement and smoothing of the contours with Savitzky-Golay filter in the remaining regions of interest. The potential of auto-LSN to diagnose advanced fibrosis was measured with the area under curve (AUC), and compared with LSN, measured with the commercially available software by a radiologist (8 years of experience). The Delong's test was used for the comparison of AUCs.
Results or Findings: One hundred and two patients were included (88 males, age 72y ± 9.89),75 (74%) in the advanced fibrosis group (Metavir F3/F4). Auto-LSN and LSN scores were 2.64 (± 0.47) and 3.50 (± 0.86) respectively in the advanced fibrosis group, 2.20 (± 0.51) and 3.00 (± 0.67) in the other. The AUC of auto-LSN and LSN were 76% and 68%, respectively (p = 0.18).
Conclusion: The performance of auto-LSN and LSN to diagnose advanced fibrosis were similar. Auto-LSN has the advantage to be fully automated.
Limitations: Monocentric retrospective study.
Funding for this study: This work is supported by a public grant overseen by the French National Research Agency (ANR) as part of the Investments for the Future programme (PIA) under grant agreement No. ANR-21-RHUS-01. This work is also funded by Guerbet.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable for this study.
7 min
Assessing the accuracy of an AI-based coronary artery calcium score algorithm on non-gated chest CT images with varying slice thicknesses
Dan Mu, Nanjing / China
Author Block: D. Mu1, K. Yin1, W. Chen1, X. Chen2, B. Zhang1; 1Nanjing/CN, 2Shanghai/CN
Purpose: This study aimed to evaluate the accuracy of an artificial intelligence-based coronary artery calcium score (AI-CACS) algorithm when applied to non-gated chest computed tomography (CT) images with varying slice width thickness.
Methods or Background: A total of 112 patients who underwent both chest CT and simultaneous electrocardiogram (ECG)-gated non-contrast enhanced cardiac CT were prospectively enrolled. Different image thicknesses (1 mm, 3 mm, and 5 mm) were reconstructed from the same chest CT scan. The coronary artery calcium score (CACS) was obtained semi-automatically from ECG-gated cardiac CT scans using a conventional CAD method, serving as the reference (ECG-CACS). An AI-based algorithm was developed to automatically calculate CACS from non-gated chest CT images (AI-CACS). Agreement and correlation were assessed using Bland-Altman analysis and Spearman correlation coefficients. Risk stratification was also performed and compared.
Results or Findings: AI-CACS demonstrated strong correlations with ECG-CACS for the three different slice thicknesses (1 mm: 0.973, 3 mm: 0.941, 5 mm: 0.834; all p < 0.001). AI-CACS with a 1 mm slice thickness showed no statistically significant difference compared to ECG-CACS (p=0.085). The Bland-Altman plot revealed mean differences of -6.5, 15.4, and 53.1 for the AI-CACS 1 mm, 3 mm, and 5 mm groups, respectively, with 95% limits of agreement of -95.0 to 81.9, -96.6 to 127.4, and -187.8 to 294.0, respectively. Agreement of risk categories for CACS was measured by kappa (κ) values (AI-CACS-1mm: 0.868; AI-CACS-3mm: 0.772; AI-CACS-5 mm: 0.412; all p < 0.001), and the concordance rate was 91%, 84.8%, and 62.5%, respectively.
Conclusion: The AI-based algorithm proved to be feasible for calculating CACS from chest CT scans, with images having a 1mm slice width thickness yielding the best results.
Limitations: Only slice thicknesses of 1 mm, 3 mm, and 5 mm were evaluated in this study. A larger multi-centre, multi-vendor cohort study shall be conducted.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the IRB with approval number: 2022-547-01.
7 min
Deep learning decision-making research software prototype for automated inner ear classification
Andras Kedves, Innsbruck / Austria
Author Block: A. Kedves1, A. A. Almasri2, S. Sugarova3, A. Alsanosi4, F. Almuhawas4, L. Hofmeyr5, F. Wagner6, K. Sriperumbudur1, A. Dhanasingh1; 1Innsbruck/AT, 2Pecs/HU, 3St. Petersburg/RU, 4Riyadh/SA, 5Stellenbosch/ZA, 6Bern/CH
Purpose: The aim of this study was to create an efficient DICOM viewer program, that automatically crops the inner ear, and classifies inner ear malformations (IEM), based on computed tomography (CT).
Methods or Background: Retrospectively we evaluated 2053 patients from three hospitals and extracted 1200 inner ear CTs to import, crop, and the artificial intelligence (AI) to train, test, and validate. An automated cropping algorithm based on K-means clustering was created to crop the inner ear volume, along with a simple graphical user interface (GUI). Using the crops as an input, we created a deep learning convolution neural network (DL CNN) (5-fold cross-validation) to determine whether the inner ear anatomy is abnormal or normal (data equally distributed). Abnormal anatomy consists of cochlear hypoplasia, ossification, incomplete partition type I, incomplete partition type III, and common cavity (data equally distributed) were selected. Both the cropping tool and the AI model were validated.
Results or Findings: We developed an efficient research software prototype that can read CT files and crops a volume that contains the inner ear. Based on that volume of interest, the AI model makes the classification. The cropping was 92.25% accurate. The area under the curve (AUC) is 0.86 (95% CI: 0.81-0.91) on DL. Accuracy, precision, recall, and F1 scores are 0.812, 0.791, 0.8, and 0.766, respectively.
Conclusion: We present a fully automatised workflow of software development and validation tool. Our solution could provide good diagnostic accuracy during risk stratification; however, must be supervised by the decision-maker.
Limitations: The most important limitation to make the AI model more robust is the number of samples available at the time of this 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: Approved bythe independent ethics committee of three hospitals IRB Nos. 22/0084/IRB, 23_001/IRB, and S_23_001/IRB, respectively.
7 min
Deep learning for estimating pulmonary nodule malignancy risk: how much data does AI need to reach radiologist level performance?
Bogdan Obreja, Nijmegen / Netherlands
Author Block: B. Obreja1, K. V. Venkadesh1, W. Hendrix1, Z. Saghir2, M. Prokop1, C. Jacobs1; 1Nijmegen/NL, 2Hellerup/DK
Purpose: Deep learning algorithms require large training datasets to achieve optimal performance. For many AI tasks, it is unclear whether algorithm performance would improve further if more training data was added. The aim of this study is to quantify the number of CT training samples required to achieve radiologist-level performance for a deep learning AI algorithm that estimates pulmonary nodule malignancy risk.
Methods or Background: For estimating pulmonary nodule malignancy risk, we used the NLST dataset (malignant nodules:1249, benign nodules:14828) to train a deep learning algorithm. The dataset was split: 80% training and 20% internal validation. The algorithm was trained on random subsets of the training set with subset sizes ranging from 10% to 100%, with a class distribution of malignant≈7.77% and benign≈92.23%. The trained AI algorithms were validated on a size-matched cancer-enriched cohort (malignant:59, benign:118) from DLCST. The performance was compared against a group of 11 clinicians that also scored the test set, which included 4 thoracic radiologists.
Results or Findings: Using training data subsets of 10%, 20%, and 30%, the AI achieved AUC values of 0.74 (95%CI:0.67-0.82), 0.79 (95%CI:0.72-0.85), and 0.81 (95%CI:0.74-0.87) respectively. When the training data set size reached 60% (malignant:602, benign:7112), the performance saturated, reaching an AUC of 0.82 (95%CI:0.75-0.88). This was comparable to the average AUC of all clinicians (0.82,95%CI:0.77-0.86,p>0.99) and of the four thoracic radiologists (0.82,95%CI:0.74-0.89,p>0.99).
Conclusion: The AI was able to reach the level of an experienced thoracic radiologist when it was trained on 7714 nodules (malignant:602) from the NLST dataset. These findings have potential implications for the allocation of resources in developing deep learning algorithms for lung cancer medical imaging diagnostics.
Limitations: The generalizability of these findings is constrained by heterogeneity and geographical limitations of the datasets used in this study.
Funding for this study: Public private consortium with funding from NWO, Dutch Ministry of Economic Affairs, and MeVis Medical Solutions, Bremen, Germany.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study included data collected from the National Lung Screening Trial (NLST) and the Danish Lung Cancer Screening Trial (DLCST). For NLST, IRB approval was obtained at all 33 participating sites and all participants provided informed consent. For DLCST, the Ethics Committee of Copenhagen County approved the study, and informed consent was obtained from all participants.
7 min
SALT: Softmax for Arbitrary Label Trees
Giulia Baldini, Essen / Germany
Author Block: G. Baldini1, S. Koitka1, C. M. Friedrich2, J. Haubold1, B. M. Schaarschmidt1, M. Forsting1, F. Nensa1, R. Hosch1; 1Essen/DE, 2Dortmund/DE
Purpose: Segmentation networks treat anatomical structures as isolated entities, neglecting their inherent hierarchical relationships. We aimed to develop Softmax for Arbitrary Label Trees (SALT) that leverages these properties to enhance segmentation speed and interpretability.
Methods or Background: This study proposes a segmentation method for CT-imaging that employs conditional probabilities to model the hierarchical structure of anatomical landmarks (such as, the lungs can be split in left/right and in upper/middle/lower lobe). This study utilizes 900 body region segmentations (883 patients) of the SAROS dataset from The Cancer Imaging Archive (TCIA). The TotalSegmentator was used to generate additional segmentations for a total of 117 labels. SALT was trained on 600 CTs, and 150 CTs were used for validation and testing. The model was evaluated on SAROS and on the TCIA dataset LCTSC using the Dice-Similarity-Coefficient (DSC).
Results or Findings: On the SAROS test set, the model obtained a DSC of 0.99 for abdominal and thoracic cavities, 0.98 for mediastinum, bones, and pericardium, 0.97 for muscles, 0.93 for subcutaneous tissue, 0.86 for brain, and 0.81 for spinal cord. On the LCTSC dataset, the model exhibited a DSC of 0.94 for right lung, 0.91 for left lung, 0.93 for both lungs, 0.89 for pericardium, 0.83 for spinal cord. Furthermore, SALT demonstrated remarkable computational efficiency, being capable of segmenting a whole-body CT in 20 seconds. This feature allows for its integration into clinical workflows, as a full-body segmentation could be automatically and efficiently computed whenever a CT scan is performed.
Conclusion: SALT used the hierarchical structures inherent in the human body to achieve high-quality segmentations while delivering exceptional speed. Additionally, this method also allows training using multiple, incompatible datasets.
Limitations: SALT was only evaluated on two TCIA datasets and its performance should be assessed on more datasets and conditions.
Funding for this study: This study did not receive external funding.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Ethical approval was not required for this study, as it exclusively utilizes data already publicly available on The Cancer Imaging Archive (TCIA).
7 min
Validation of an artificial intelligence software for automatic pulmonary nodule volumetry using micro-CT determined ground truth nodule volumes
Louise D'hondt, Ghent / Belgium
Author Block: L. D'hondt1, P-J. Kellens1, K. Torfs2, H. Bosmans2, A. Snoeckx3, K. Bacher1; 1Ghent/BE, 2Leuven/BE, 3Antwerp/BE
Purpose: Validation of pulmonary nodule volumetry in clinically available software for automatic nodule detection and volumetry is currently either underrepresented or based on a consensus-driven ground truth in patient images, introducing uncertainties due to intrinsic structural differences between readers. Purpose of the study was to validate the nodule volumetry of a computer-aided detection (CAD) software using objective micro-CT determined ground truth nodule volumes.
Methods or Background: Eighteen 3D-printed solid pulmonary nodules, including six diameters and three morphology classes, were subjected to high-resolution μCT scanning to establish objective ground truth volumes. The anthropomorphic Lungman phantom, containing the nodules, was scanned using a 256-slice CT scanner at three CTDIvol-values (6.04, 1.54, 0.20 mGy), and subsequently reconstructed with both iterative and deep learning image reconstruction, along with either soft or hard kernels. Volumetric accuracy of a commercially available automatic volumetry software (AVIEW LCS+) was assessed through multiple linear regression, identifying which predictors (reconstruction algorithm, kernel, dose, morphology, and diameter) significantly influence the outcome (% error).
Results or Findings: Volumes of nodules larger than six mm in diameter were accurate within 10% of their ground truth volume. Accuracy of volume measurements was significantly influenced by smaller and irregular morphologies (p<0.001). Notably, variation in reconstruction algorithm exerted no significant influence (p>0.05). Radiation dose and reconstruction kernel emerged as crucial parameters for accuracy. However, regression analysis showed diminished impact of the latter through significant interaction with nodule characteristics, resulting in measured volumes closer to the real volumes. Considerable discrepancies were observed between objective pulmonary nodule volumes and volumes determined through consensus readings.
Conclusion: Robustness of the volumetry software for variations in CT acquisition parameters in a phantom makes it a valuable candidate application for diverse imaging protocols across multiple centers.
Limitations: This is a phantom study.
Funding for this study: Funding was provided by the FWO “Kom op tegen Kanker”-project for lung cancer screening research in Belgium. (Project number: G0B1922N).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: No ethics committee approval was needed, since this study used phantom images.
7 min
AI-assisted detection of abdominal pathologies in chest CT scans
Mikhail Belyaev, London / United Kingdom
Author Block: E. Petrash, M. Dugova, R. Gareeva, E. Kochina, A. Shevtsov, V. Samokhin, F. Yaushev, M. Belyaev; London/UK
Purpose: Amid intensified focus on chest pathologies, abdominal findings on chest CT scans may be overlooked. Detecting non-alcoholic fatty liver disease (NAFLD) is crucial due to its hepatocellular carcinoma risk. Early identification of liver, kidney, and adrenal incidentalomas aids in timely cancer interventions. Swift urolithiasis recognition prevents ureteral complications. Identifying aortic aneurysms is vital due to its high mortality. Our research assesses an AI solution's efficacy in enhancing these detection rates on chest CT.
Methods or Background: The retrospective study used data from the national lung screening trial, consisting of 2408 chest CTs from primary patients. We excluded 130 cases with a pathology below the diaphragm mentioned in the report. The remaining 2278 CTs without described abdominal findings were auto-analysed by a comprehensive AI product, AUCT-Abdomen, which detects adrenal mass, NAFLD, liver & kidney lesions, urolithiasis, and abdominal aortic dilatation. CTs with AI findings were then independently reviewed by two radiologists with 12 and 14 years of experience.
Results or Findings: AI identified 270 (11.2%) previously unreported patients on top of 130 (5.3%) reported initially. AI findings include 125 adrenal mass, 21 NAFLD, 68 liver lesions, 17 kidney lesions, 5 urolithiasis, 20 aortic dilatation, and 5 aortic aneurysm. AI results were false positives in 17 (0.7%) cases.
Conclusion: Integrating AI into the radiological evaluation of chest CT scans may ensure the thorough detection of abdominal incidental findings, including adrenal mass, hepatic steatosis, liver and kidney lesions, urolithiasis, aortic dilatation and aneurysm. Our research may highlight a potential deprioritisation of findings peripheral to the chest and the potential of AI in bridging this diagnostic gap. The potential benefits from deploying AI products for incidental abdominal findings may be exceptionally high for departments with strictly limited turnaround time.
Limitations: No limitations were identified for this study.
Funding for this study: Funding for this study was provided by AUMI AI Limited.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: The study is retrospective so no ethical approval was sought.
7 min
Fully automatic AI-driven assessment in coronary CT angiography for intermediate stenosis: a comparative study with quantitative coronary angiography and fractional flow reserve
Jung-In Jo, Seoul / Korea, Republic of
Author Block: J-I. Jo, H. J. J. Koo, J-W. Kang, D. H. Yang; Seoul/KR
Purpose: Limited data exists on direct comparison of AI-driven automatic coronary stenosis assessment in coronary CT angiography (CCTA) to quantitative coronary angiography (QCA). This study aims to compare AI-based coronary stenosis evaluation in CCTA with its quantitative counterpart of coronary angiography and invasive fractional flow reserve (FFR).
Methods or Background: In this single-center retrospective study, 215 intermediate coronary lesions, with QCA diameter stenosis between 20% and 80%, were assessed from 195 symptomatic patients (mean age 61 years, 149 men, 585 coronary arteries). For stenosis quantification in CCTA, an AI-driven research prototype (Siemens Healthineers, Germany) was used (AI-CCTA). Diagnostic accuracy of AI-CCTA on per-vessel basis was assessed, using invasive coronary angiography stenosis grading (with > 50% stenosis) or invasive FFR (< 0.80) as reference standards. AI-driven diameter stenosis was then correlated with QCA results and expert manual measurements.
Results or Findings: Among 585 coronary arteries, disease prevalence as determined by invasive angiography (≥ 50%) was 46.5%. AI-CCTA showed sensitivity of 71.7%, specificity of 89.8%, positive predictive value of 85.9%, negative predictive value of 78.5%, and area under the curve (AUC) of 0.81. For 215 intermediate lesions assessed using QCA and FFR, diagnostic performance of AI-CCTA was moderate, with AUC of 0.63 for both QCA and FFR. In measuring degree of stenosis, AI-CCTA demonstrated a moderate correlation with QCA (r = 0.42, p < 0.001), which was notably better than results from manual quantification versus QCA (r = 0.26, p=0.001).
Conclusion: The AI-powered automated CCTA analysis showed promising results when compared to invasive angiography. While AI-CCTA demonstrated a moderate relationship with QCA in intermediate coronary stenosis lesions, its results appeared to surpass those of manual evaluations.
Limitations: Individuals with stents or a history of coronary-artery bypass grafting were excluded from analysis.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval from the Institutional Review Board of Asan Medical Center, Seoul, Korea
7 min
Artificial intelligence-based non-contrast-enhanced CT images for diagnosis of hepatic lesions: a multicentre study
Zhuangxuan Ma, Shanghai / China
Author Block: Z. Ma, L. Jin, M. Li; Shanghai/CN
Purpose: We aimed to realise the identification and classification diagnosis of intrahepatic space-occupying lesions on non-contrast enhanced CT (NCCT).
Methods or Background: In this retrospective study, patients who suspected with space-occupying lesions in liver undergoing both NCCT and enhanced CT/MRI multi-phase enhancement from January 2017 to March 2023 in our hospital and from January 2020 to August 2023 in another medical center. Each liver lesions in NCCT were confirmed by enhanced CT/MRI multi-phase enhancement or pathology as the golden reference. The lesion contours of the patients in NCCT images were manually delineated by radiologists, and radiomics features were extracted in 3D Slicer. An automatic machine learning algorithm was used to screen out the most relevant radiomics features and establish a classification model for differential diagnosis to classify type of intrahepatic space-occupying lesions.
Results or Findings: A total of 252 liver lesions in 230 patients from our hospital including 79 hepatic cysts, 81 haemangiomas, 52 malignant tumours and 40 liver abscesses. A total of 33 liver lesions in 230 patients from another medical centre including 12 hepatic cysts, 8 haemangiomas, 8 malignant tumours and 5 liver abscesses. The sensitivity of the nnDetection to detect the lesion was 0.81, the AUC of classification model for lesions were 1.0 (hepatic cysts), 0.99 (hemangiomas), 0.9 8(malignant tumours) and 0.98 (liver abscesses) in our internal test dataset while the AUC were 1.0 (hepatic cysts), 0.97 (haemangiomas), 0.9 3 (malignant tumours) and 0.92 (liver abscesses) in outside test dataset.
Conclusion: Our proposed model showed better performance of the identification and classification diagnosis of intrahepatic space-occupying lesions on NCCT.
Limitations: The small sample size may lead the bias of this 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: The study was approved by Huadong hospital (reference number: 20230089).

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