{
  "abstract": "Background Right ventricular (RV) mass is a potentially important biomarker of cardiac health, offering diagnostic insights into conditions such as pulmonary hypertension and congenital heart disease. Cardiovascular magnetic resonance (CMR) is the gold standard for RV mass quantification, but manual segmentation of the RV myocardium is labour-intensive, hindering the clinical translation of this informative biomarker. Recent advances in deep learning (DL), particularly convolutional neural networks (CNNs) and Transformers, have enabled automation of the segmentation of the left ventricle blood pool and myocardium as well as the RV blood pool, but a robust too for automated RV myocardium segmentation has not yet been implemented.Methods Cine short-axis CMR datasets containing end-diastolic and end-systolic images were acquired from King’s College London/Guy's and St Thomas' NHS Foundation Trust. Manual segmentation of the RV myocardium was carried out by an experienced cardiologist to act as a ground truth. 313 subjects were utilised for training three 3D DL models: nnU-Net (CNN-based), SwinUNETR (Transformer-based) and MedNeXt (hybrid CNN-based), using five-fold cross-validation on the training set. The resulting models were evaluated on a test set of 79 subjects using three metrics: Dice Similarity Coefficient (DSC) for segmentation accuracy, 95% quantile of Hausdorff Distance (HD95) for contour accuracy, and RV mass estimation to evaluate clinical applicability. Additionally, the impact of preprocessing strategies, specifically the selection of target spacing (anisotropic vs isotropic), was also investigated. The configuration of all training networks is shown in the table 1.Abstract 6-026 Table 1Hyperparameters used during training. Both models nnUNetv2 and nnUNetv1 represent the same model variant (nn-UNet) with different resampling preprocessing, while nnUNetv1, SwinUNETR and MedNeXt are different model architectures with the same resampling preprocessing. The patch size and target spacing were adjusted, except for nnUNetv2, which is automatically configured No Model Architecture Based Preprocessing (Rule-based) Patch Size Target Spacing 1nnUNetv2 (Baseline)CNN[16, 320, 320][8.0, 1.194, 1.194]2nnUNetv1CNN[128, 128, 128][1.0, 1.0, 1.0]3SwinUNETRTF[128, 128, 128][1.0, 1.0, 1.0]4MedNeXtCNN[128, 128, 128][1.0, 1.0, 1.0]Results As shown in the table 2, the CNN-based nnUNet achieved the best results with a mean DSC of 61.12% and HD95 of 7.28 mm for the RV myocardium, outperforming the Transformer-based SwinUNETR. MedNeXt demonstrated competitive performance with potential for further optimisation. Preprocessing strategies, specifically the selection of target spacing, had a significant impact on segmentation accuracy. As illustrated in the figure 1, RV mass estimations agreed reasonably well with ground truth values, but with some failure cases and a consistent overestimation of RV mass across all models.Abstract 6-026 Table 2Overall metric statistics for RV myocardium segmentation and contour accuracy on the test set across all applied models, with mean and standard deviation. Best results are in bold Model RV myocardium DSC (%) HD95 (mm) nnUNetv158.24 ± 12.988.696 ± 4.943nnUNetv2 (Baseline)61.12 ± 13.157.275 ± 3.936SwinUNETR56.58 ± 12.569.791 ± 5.642MedNeXt61.20 ± 12.777.684 ± 4.226Abstract 6-026 Figure 1Conclusion This study presents the first automated DL-enabled pipeline for RV myocardium segmentation. The results demonstrate the feasibility of DL techniques for large-scale, clinically applicable RV mass estimation. While CNN-based architectures outperformed Transformer-based models, further research is required to enhance generalisability and scalability. This pipeline provides a robust foundation for advancing automated cardiac imaging in clinical practice.",
  "authors": [
    {
      "affiliations": [
        "King’s College London, London, UK"
      ],
      "name": "Guokai Zeng"
    },
    {
      "affiliations": [
        "King’s College London, London, UK"
      ],
      "name": "Andrew King"
    },
    {
      "affiliations": [
        "King’s College London, London, UK"
      ],
      "name": "Pier-Giorgio Masci"
    },
    {
      "affiliations": [
        "DTU - Technical University of Denmark, Denmark"
      ],
      "name": "Tareen Dawood"
    },
    {
      "affiliations": [
        "HeartFlow, Inc"
      ],
      "name": "Esther Puyol Anton"
    }
  ],
  "title": "6-026 Deep learning-based assessment of cardiac right ventricular mass from MR imaging",
  "uid": "7fe40e59-5771-5aab-bb26-16e820efbaf5"
}