Deep Learning for Ventricular Arrhythmia Prediction Using Fibrosis Segmentations on Cardiac MRI Data

Florence E. van Lieshout, Roel C. Klein, Maarten Z. Kolk, Kylian van Geijtenbeek, Romy Vos, Samuel Ruipérez-Campillo, Ruibin Feng, Brototo Deb, Prasanth Ganesan, Reinoud Knops, Ivana Isgum, Sanjiv M. Narayan, Erik J. Bekkers, Bob D. de Vos, Fleur V. Y. TjongCinC 2022 · OpenReview

Introduction

Sudden cardiac death (SCD) remains a significant global health challenge, frequently precipitated by ventricular arrhythmias (VA). Currently, the primary preventive measure is the Implantable Cardioverter Defibrillator (ICD). However, clinical guidelines for ICD implantation—largely based on Left Ventricular Ejection Fraction (LVEF)—are imprecise. Statistics indicate that up to 83% of recipients do not require a life-saving shock within the first four years, while many high-risk patients are overlooked.

Medical research suggests that the geometry and extent of myocardial fibrosis (scarring), visible on Late Gadolinium Enhanced (LGE) MRI, are potent predictors of VA. This paper, presented at Computing in Cardiology 2022, proposes a two-stage deep learning pipeline to automate fibrosis segmentation and improve risk stratification using Graph Neural Networks (GNNs).

Methods

The proposed framework consists of a multi-step architectural pipeline designed to leverage anatomical priors for enhanced segmentation accuracy and geometric analysis for prediction.

  1. Two-Stage Segmentation Pipeline:
  2. M-Unet: A standard 2D U-Net is utilized to segment the myocardium (heart muscle) from raw MRI slices.
  3. F-Unet: A subsequent U-Net segments the fibrosis. This model incorporates the myocardium probability map from the first stage as a "location prior," focusing the search for scarring within the heart muscle. The study evaluated 2D, 2.5D (employing 3D convolutions in the bottleneck), and full 3D variants.
  4. ICD Therapy Prediction (GNN): To capture the complex spatial distribution of fibrosis, the authors employed an $E(n)$ equivariant Graph Neural Network. The heart is represented as a graph by sampling 100 fibrosis voxels and 500 myocardium voxels as nodes, with edges connecting each node to its 25 nearest neighbors. This allows the model to learn structural patterns and spatial relationships rather than relying solely on volumetric measurements.
The complete architectural pipeline, from raw MRI to ICD therapy prediction.

The study utilized the DEEP RISK dataset, which includes data from 1,261 ICD patients. Segmentation models were trained on a subset of 117 expert-annotated patients, while the prediction task was evaluated on 515 patients with sufficient imaging data.

Results

The experimental results highlight the efficacy of incorporating anatomical context and geometric representation over traditional clinical metrics.

  • Segmentation Performance: The 2.5D F-Unet emerged as the optimal configuration, achieving a Dice Similarity Coefficient (DSC) of 0.769. This outperforms previous benchmarks (approx. 0.71). The authors suggest the 2.5D approach effectively captures inter-slice context while remaining robust to the inconsistent voxel spacing typical of cardiac MRI.
  • Prediction Accuracy: The myo-GNN model reached an AUC of 0.60, significantly outperforming traditional clinical markers. For comparison, logistic regression models based on current standards like LVEF and the mere presence of LGE yielded AUCs of 0.50 and 0.55, respectively. This suggests that the spatial arrangement of scar tissue is a more critical predictor of arrhythmia than total volume.
Visual examples of segmentation results across three different patients, comparing ground truth with various model predictions.

Discussion and Conclusion

The integration of automated segmentation with graph-based geometric analysis represents a significant step toward more precise SCD risk assessment. By using the myocardium as a spatial prior, the segmentation models achieved state-of-the-art performance. Furthermore, the GNN's superior performance over volumetric baselines underscores the importance of cardiac scar topology in arrhythmogenesis.

While an AUC of 0.60 indicates substantial improvement over current standards, it is not yet sufficient for autonomous clinical decision-making. Future research will focus on multimodal frameworks that integrate these geometric MRI features with other clinical data, such as ECG signals and patient demographics, to further enhance predictive accuracy.

Summary updated

BibTeX
@inproceedings{lieshout2022deep,
  author = {Florence E. van Lieshout and Roel C. Klein and Maarten Z. Kolk and Kylian van Geijtenbeek and Romy Vos and Samuel Ruipérez-Campillo and Ruibin Feng and Brototo Deb and Prasanth Ganesan and Reinoud Knops and Ivana Isgum and Sanjiv M. Narayan and Erik J. Bekkers and Bob D. de Vos and Fleur V. Y. Tjong},
  title = {{Deep Learning for Ventricular Arrhythmia Prediction Using Fibrosis Segmentations on Cardiac MRI Data}},
  booktitle = {CinC},
  year = {2022},
  url = {https://openreview.net/forum?id=IGNRqE0pwO},
}
Figure