Research Presentation Session: Artificial Intelligence & Machine Learning & Imaging Informatics
RPS 105 - Automatic segmentation techniques
February 28, 08:00 - 09:00 CET
7 min
Deep learning for segmentation and classification of cardiac implantable electronic devices on chest x-rays
Felix Busch, Munich / Germany
- Author Block: F. Busch1, A. Zhukov1, P. Suwalski1, S. Niehues1, D. Poddubnyy1, M. Makowski2, K. K. Bressem1, L. C. Adams2; 1Berlin/DE, 2Munich/DEPurpose: The accurate classification of cardiac implantable electronic devices (CIEDs) on chest x-rays is crucial for effective patient care. The aim of this study was to create an open-access deep learning algorithm capable of both segmenting and classifying CIEDs on DICOM as well as smartphone-acquired images for bedside use.Methods or Background: This retrospective study included patients with implantable pacemakers, cardioverter defibrillators, cardiac resynchronisation therapy devices, and cardiac monitors who had undergone anterior-posterior or posterior-anterior chest radiography from January 2012 to January 2022 at Charité – University Medicine Berlin. Utilising a U-Net architecture with a ResNet-50 backbone, we developed a model to segment and classify CIEDs based on their manufacturer and model, using both DICOM and smartphone images. Performance metrics included the Dice coefficient for the segmentation model on the validation set (70-30 training/validation set split) and balanced accuracy for manufacturer and model classification on the test set (70-20-10 training/validation/test set split).Results or Findings: The study encompassed 897 patients with 2,322 unique chest radiographs featuring 25 CIED models from six manufacturers. To prevent misclassification of models less represented or not included in the training data, an "other" category was implemented. Additionally, 11,072 images were captured using five different smartphones. The segmentation algorithm attained an average Dice coefficient of
- 936 (interquartile range: 0.068), while the classification model achieved an overall accuracy of 0.927 (95% confidence interval (CI): 0.890-0.965) for manufacturer and 0.847 (95% CI: 0.799-0.888) for model classification. Conclusion: We present a publicly accessible deep learning framework for the high-accuracy segmentation and classification of CIEDs on chest x-rays. Notably, this research introduces the first classification algorithm specifically designed for accurate CIED model identification based on both DICOM and smartphone images.Limitations: The retrospective design of the study and the unequal representation of CIEDs were identified as limitations.Funding for this study: No funding was received for this study.Has your study been approved by an ethics committee? YesEthics committee - additional information: This study was approved by an ethics committe; IRB-approval number: EA4/042/
7 min
Advancing total tumour volume estimation in colorectal liver metastases: development and evaluation of a self-learning auto-segmentation model
Inez Margaretha Verpalen, Amsterdam / Netherlands
- Author Block: J. I. Bereska1, M. Zeeuw1, L. Wagenaar1, M. G. Besselink1, H. Marquering1, J. Stoker1, Å. Fretland2, G. Kazemier1, I. M. Verpalen1; 1Amsterdam/NL, 2Oslo/NOPurpose: Total tumour volume (TTV) assessments have been shown to be prognostic of overall and recurrence-free survival in patients with colorectal cancer liver metastases (CRLM). However, the labor-intensive nature of these assessments has hampered their clinical adoption. This study aimed to develop an auto-segmentation model for CRLM on contrast-enhanced portal venous phase CT scans to facilitate the clinical adoption of TTV assessments.Methods or Background: We developed a self-learning-based segmentation model to segment CRLM using 760 portal venous phase CTs (CT-PVP) of 363 patients with 13,739 CRLM from the Amsterdam University Medical Centre. We used a self-learning setup in which we first trained a teacher model on 99 manually segmented CT-PVPs segmented by three radiologists and combined using the Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm. The teacher model was then used to segment the remaining 661 CT-PVPs for training the student model. We used Intraclass Correlation Coefficient (ICC) to compare the TTV obtained from the student model's segmentations against that obtained from the STAPLE-combined radiologist's segmentations.Results or Findings: We evaluated the student model in an external test set of 50 CT-PVPs from 35 patients with 72 CRLM from the Oslo University hospital. The student model reached a DICE similarity score of
- 83 for segmenting CRLM. There was no significant difference between the student model's DICE scores and interrater DICE scores. The ICC between the student model's and the STAPLE-combined TTV was 0.97, signifying near perfect agreement. Conclusion: Segmentation models can provide accurate and efficient assessments of TTV in CRLM patients.Limitations: Our study's limitations include its retrospective design, lack of global data, and an external test cohort that differs from the training set, underlining the need for prospective, internationally diverse studies for more robust validation.Funding for this study: This study was funded by the KWF (project number 14002).Has your study been approved by an ethics committee? YesEthics committee - additional information: The Medical Ethics Review Committee of the Amsterdam UMC, the Regional Ethical Committee of South Eastern Norway, and the Data Protection Officer of Oslo University Hospital approved this study protocol.
7 min
A deep learning-based pipeline for cervical spinal cord segmentation and labelling on heterogeneous T1w brain images
Ricardo Magalhaes, Braga / Portugal
- Author Block: R. Magalhaes, A. Liseune, S. van Eyndhoven, T. Billiet, N. Barros, D. Smeets, D. M. Sima; Leuven/BEPurpose: The aim of this study was to develop a robust and automated deep learning-based method for cervical spinal cord measurements on T1w brain images.Methods or Background: Measuring spinal cord (SC) cross-sectional area (CSA) is valuable for monitoring multiple sclerosis (MS), but challenging in daily clinical practice. We propose a fully automated processing pipeline that performs this measurement in mere minutes and is robust for a wide range of imaging protocols. The pipeline was developed using T1w brain scans from MS patients, with ground truth masks generated using an in-house semi-automated pipeline that ensures SC coverage and segmentation quality (187 training, 44 validation). Starting from an input T1w image, the pipeline uses icobrain to perform neck cropping and subsequently applies three cascaded U-net deep learning models that respectively segment, smooth and label the spinal cord, from which the measurements are derived. We report performance on an independent data set containing 10 MS subjects with 53 scans from different scanners.Results or Findings: Dice scores for the segmentation of the full SC and for labelling vertebrae C1 to C4 were
- 89 and 0.85, 0.87, 0.85 and 0.83, respectively. Intra-scanner measurement reproducibility on the test set was high, with an average relative CSA error of 1.5% (intrascanner) and 4.6% (interscanner). Conclusion: Trained on a heterogeneous set of T1w brain scans, the pipeline enables reliable and accurate quantification of cervical SC using standard brain scans, extending icobrain software's capabilities.Limitations: The method requires brain images covering at least a portion of the cervical spine.Funding for this study: This study is partly funded by Flanders Innovation & Entrepreneurship (VLAIO) project HeKDiscoMS (HBC.
- 0500) and by the CLAIMS project, supported by the Innovative Health Initiative Joint Undertaking (JU) under grant agreement No 101112153. The JU is supported by the European Union's Horizon Europe research and innovation programme and COCIR, EFPIA, EuropaBio, MedTech Europe, Vaccines Europe, AB Science SA and icometrix NV. Has your study been approved by an ethics committee? YesEthics committee - additional information: The test data set was acquired from 10 MS patients who participated in a study at the University Hospital Brussels, Belgium. The study was approved by the local ethics committee and all patients signed informed consent forms.Development data comes from subjects for which icometrix analysed MR scans as part of clinical practice who had agreed to allow icometrix to use an anonymised version of the already analysed MR images for post-market research purposes. Icometrix processes personal data received from the hospitals in conformity with the applicable data protection and privacy legislation.
7 min
Multi-organ CT-based automatic segmentation via semi-supervised learning
Alejandro Vergara, Valencia / Spain
- Author Block: A. Vergara, A. Jimenez-Pastor, C. Kerckhaert, A. Alberich-Bayarri; Valencia/ESPurpose: To overcome the scarcity of manually annotated data sets for multi-organ segmentation, we propose a novel approach that combines fragmented data sets to train a single model capable of performing multiorgan segmentation.Methods or Background: Four distinct data sets were collected: AbdomenCT1k (liver, kidneys, pancreas, spleen), CT-org (lungs, bladder), VerSe (spine), and CTPelvic1K (hips, sacrum), collecting a total of 1543 cases. Each data set was used to train a 2D U-Net with deep supervision. These submodels generated pseudo-labels for data sources that were not included in their training, resulting in a combination of both original strong labels and soft labels. Finally, a unified model was trained using all data sets and labels.Results or Findings: The submodels achieved a mean Dice Score Coefficient (DSC) of
- 91. The final model improved the DSC for each structure by a mean of 0.15, attaining a maximum DSC of 0.98 for liver and a minimum of 0.78 for bladder. Conclusion: Our study introduces an innovative method for training a single model using diverse data sources, leveraging a pseudo-label semi-supervised strategy to achieve robust multiorgan segmentation. This approach enables the generation of a larger annotated data set from smaller, specialised ones lacking all desired labels. Additionally, employing a unified model, as opposed to separate models for each data source, offers advantages in terms of reduced inference time and resource efficiency.Limitations: The accuracy of the pseudo-labels used for training the final model is crucial. Gross errors or inaccuracies in these labels could propagate through the subsequent stages of training, affecting the final model's performance.Funding for this study: The IMAS project (High Sensitivity and Low Dose Molecular Imaging) was funded by the Spanish Ministry of Science and Innovation and European Funds.Has your study been approved by an ethics committee? Not applicableEthics committee - additional information: No information provided by the submitter.
7 min
ACMA-net and unscented Kalman filter-based accurate coronary artery segmentation: an application of deep learning to computed tomography angiography image
Bao Li, Beijing / China
- Author Block: B. Li, C. Wen, H. Sun, W. Wang, J. Liu, Y. Liu; Beijing/CNPurpose: Accurate coronary computed tomography angiography (CCTA) image segmentation is a prerequisite for high-precision reconstruction of three-dimensional (3D) coronary artery models, which can visually demonstrate stenotic lesion information and develop treatment plans. However, due to the complex structure and small size of coronary arteries, and the interference in image acquisition, the 3D models reconstructed by existing segmentation technology present insufficient precision. This results in the 3D coronary artery models being noisy and prone to disconnection. To overcome the challenges in accurate segmentation of coronary arteries, this study proposes a deep learning-based two-stage algorithm.Methods or Background: In the first stage, we added an atrous convolution feature fusion module (ACFFM) and a multiaxis attention module (MAM) to 3D U-Net, called ACMA-Net, to enhance the feature expression ability of the network and effect the preliminary segmentation of coronary arteries. CCTA images of 323 patients were clinically collected to train the network. In the second stage, the preliminary segmentation results were skeletonised and endpoint detection was performed. The regions of coronary artery disconnection were determined by finding minimum distance between the main trunk branch and the endpoint of each disconnected branch. The disconnected skeleton was repaired after reconnection by the unscented Kalman filter (UKF) algorithm.Results or Findings: We evaluated the proposed method on the constructed test set of 50 patients, and the Dice and Jaccard scores were
- 940 and 0.888, respectively, outperforming existing deep learning methods. Conclusion: This study proposed a coronary segmentation method that effectively reduces the phenomenon of coronary disconnection and improves the accuracy and continuity of coronary segmentation using a small-size data set. This provides excellent technical support for the patient-specific 3D demonstration of coronary arteries.Limitations: More CTA images will be clinically collected to further validate the segmentation method.Funding for this study: Funding was provided by the National Key Research and Development Program of China (Grant No. 2021YFA1000201), and the National Natural Science Foundation of China (Grant No. 12202022, 11832003, 32271361, 12102014).Has your study been approved by an ethics committee? YesEthics committee - additional information: The Ethics Committee of Peking University People's Hospital approved this study.
7 min
Deep learning-based bowel automatic segmentation and visualisation of Crohn's disease using multilabelled continuous MRE images
Li Huang, Guangzhou / China
- Author Block: L. Huang1, Z. Zhong2, B. Huang2, S-T. Feng1, X. Li1; 1Guangzhou/CN, 2Shenzhen/CNPurpose: Recognition of bowel segments from magnetic resonance enterography (MRE) images is quite challenging and time-consuming due to unclear boundary, shape, size, and appearance variations. We established a publicly available whole bowel segments MR data set with benchmark results and visualisation.Methods or Background: We retrospectively collected T2-weighted coronal MRE data from 70 patients with Crohn's disease (CD). The bowel images per patient were divided into ten segments (stomach, duodenum, small intestine, appendix, caecum, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum), with fine pixel level annotations labelled by two experienced radiologists. Then, nnU-Net model, a deep learning-based segmentation method that automatically configures all hyperparameters based on the data set characteristics, was employed on this data set (training set, n=56; test set, n=14). To reinforce the mutually exclusive relationship between tags, a topological interaction loss function was utilised. The segmentation algorithm was assessed using the dice similarity coefficient (DSC).Results or Findings: Evaluating the performance of bowel segmentation, the mean DSC in the test set was
- 778. Our nnU-Net method in segmenting digestive tract can achieve DSC of 0.963 ± 0.042 in stomach, 0.886 ± 0.049 in duodenum, 0.936 ± 0.024 in small intestine, 0.378 ± 0.441 in appendix, 0.598 ± 0.294 in cecum, 0.825 ± 0.131 in ascending colon, 0.819 ± 0.231 in transverse colon, 0.819 ± 0.186 in descending colon, 0.801 ± 0.220 in sigmoid colon and 0.859 ± 0.130 in rectum, respectively. Segmentation results with predicted bowel boundary can be shown by two- or three-dimensional visual representation. Conclusion: We presented a new data set containing labels for all digestive tract segments on MRE images. Accurate deep learning-based bowel automatic segmentation and visualisation of CD can facilitate the application of artificial intelligence in CD.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: No information provided by the submitter.
7 min
Replicability of outcome prediction across IPF patient cohorts based on machine learning features learned without supervision
Jeanny Pan, Vienna / Austria
- Author Block: J. Pan, J. Hofmanninger, K-H. Nenning, F. Prayer, S. Röhrich, N. Sverzellati, V. Poletti, H. Prosch, G. Langs; Vienna/ATPurpose: Idiopathic pulmonary fibrosis (IPF) is most common interstitial lung disease. This study validated the transferability of an outcome prediction method based on previously identified disease patterns on a new cohort.Methods or Background: Four lung CT patterns associated with disease progression had been previously identified from 74 IPF patients using unsupervised machine learning. We studied a different multicentre cohort (various manufacturers, slice thickness and reconstruction kernels) of 164 patients to investigate the transferability of the progression patterns. We tested outcome prediction based on patterns in a single CT scan, and based on additional pattern changes in subsequent scan pairs. In both experiments, patients were clustered based on similarities in their progression pattern profiles, and Kaplan-Meier survival curves were analysed for each cluster to test if the outcome was significantly different.Results or Findings: Of the 164 patients, 59 died and 16 had transplants before the censoring date, with an average time of
- 73 weeks from the baseline scan, while for the remaining 89, it was 197.34 weeks. Predicting survival outcomes with a single scan profile yields a hazard ratio (HR) of 5.39 (p<0.01). Consistent with the results on the initial cohort, incorporating the change of pattern profiles between two scans further improved the prediction, yielding an HR of 6.03 (p <0.01). Conclusion: The replication of outcome prediction with previously identified progression markers in a new cohort of IPF patients demonstrated significant predictive value for outcome. The dynamic changes in marker profiles between scans enhanced the hazard ratio. This underscores the potential of quantitative marker profiles in disease monitoring for IPF patients. Future studies may explore the broader applicability of the method to other interstitial lung diseases.Limitations: We did not perform specific analysis to differentiate the impact of centres and manufacturers.Funding for this study: Funding for this project was received via the FWF, ONSET Project PHas your study been approved by an ethics committee? Not applicableEthics committee - additional information: No information was provided by the submitter.
This session will not be streamed, nor will it be available on-demand!
This session is accredited with 1 CME credit.
Liliana Caldeira
Cologne / Germany
Felix Busch
Munich / GermanyInez Margaretha Verpalen
Amsterdam / NetherlandsRicardo Magalhaes
Braga / PortugalAlejandro Vergara
Valencia / SpainBao Li
Beijing / ChinaLi Huang
Guangzhou / ChinaJeanny Pan
Vienna / Austria