Abstract
Whole Heart Segmentation (WHS) is a critical task in cardiology, enabling surgical planning and patient-specific modeling. This paper presents the results of the Multi-Modality Whole Heart Segmentation (MM-WHS) challenge, which benchmarked 12 state-of-the-art algorithms on a dataset of 120 clinical CT and MRI volumes. The challenge evaluated the segmentation of seven key cardiac substructures. Results indicate that while deep learning (DL) models achieve high accuracy in CT, MRI remains challenging due to image quality issues. Furthermore, traditional multi-atlas segmentation (MAS) methods demonstrate greater robustness in maintaining anatomical consistency compared to some DL approaches.
Introduction: The Challenge of Whole Heart Segmentation
Delineating the ventricles, atria, myocardium, and great vessels—collectively known as Whole Heart Segmentation—is essential for modern cardiac care. Manual segmentation is labor-intensive, often requiring up to eight hours for a single patient scan. Automating this process is complicated by significant inter-patient anatomical variability and fluctuations in image quality across different scanners and pathologies. This paper addresses these challenges by providing a standardized, multi-modality benchmark for evaluating automated WHS algorithms across both CT and MRI.
Methods
The challenge evaluated 12 international submissions, primarily categorized into two methodological frameworks:
- Deep Learning (DL): Most teams utilized Convolutional Neural Networks (CNNs), such as U-Nets and Fully Convolutional Networks (FCNs). Innovations included two-step pipelines for localization and fine segmentation, as well as multi-view approaches for 3D volume processing.
- Multi-Atlas Segmentation (MAS): This traditional approach involves registering a library of pre-labeled "atlas" images to a target scan and fusing the labels. MAS is noted for its stability and ability to preserve realistic anatomical shapes.
Datasets
The benchmark utilized a high-quality dataset of 120 3D cardiac images (60 CT and 60 MRI) acquired in real-world clinical settings. For each modality, 20 images were provided for training, while 40 were reserved for blinded evaluation. Experts labeled seven substructures: Left Ventricle (LV), Right Ventricle (RV), Left Atrium (LA), Right Atrium (RA), LV Myocardium (Myo), Ascending Aorta (AO), and Pulmonary Artery (PA).
Key Insights and Results
- Modality Performance Gap: CT segmentation generally outperformed MRI. The average Dice score for CT (~0.87) was consistently higher than for MRI (~0.82), largely due to CT's superior contrast and consistency.
- Overfitting and Generalization: Several DL models performed well on training data but failed during blinded testing, suggesting that the limited training set (20 images) led to memorization rather than anatomical learning.
- Anatomy vs. Accuracy: While DL often achieved higher overlap scores (Dice), it occasionally produced anatomically impossible artifacts. In contrast, MAS methods consistently produced plausible heart geometries.
- Substructure Difficulty: The Pulmonary Artery (PA) and Ascending Aorta (AO) proved the most difficult to segment due to their complex branching and high variability.
- Top Performers: Team "GUT" led the CT category (Dice 0.908) using a two-step CNN, while team "UB2*" led the MRI category (Dice 0.874).
Conclusion and Future Work
The primary bottleneck for automated WHS remains data scarcity. Future research should focus on incorporating anatomical constraints into neural networks, utilizing data augmentation (e.g., GANs), and exploring cross-modality learning to leverage high-quality CT data for improving MRI segmentation.
Summary updated