Abstract
We present an interactive algorithm to segment the heart chambers and epicardial surfaces, including the great vessel walls, in pediatric cardiac MRI of congenital heart disease. Accurate whole-heart segmentation is necessary to create patient-specific 3D heart models for surgical planning in the presence of complex heart defects. Anatomical variability due to congenital defects precludes fully automatic atlas-based segmentation. Our interactive segmentation method exploits expert segmentations of a small set of short-axis slice regions to automatically delineate the remaining volume using patch-based segmentation. We also investigate the potential of active learning to automatically solicit user input in areas where segmentation error is likely to be high. Validation is performed on four subjects with double outlet right ventricle, a severe congenital heart defect. We show that strategies asking the user to manually segment regions of interest within short-axis slices yield higher accuracy with less user input than those querying entire short-axis slices.
The Challenge of CHD Segmentation
Surgeons planning complex operations for pediatric Congenital Heart Disease (CHD)—such as Double Outlet Right Ventricle (DORV)—increasingly rely on patient-specific 3D models. However, the primary bottleneck is segmentation: the process of delineating the blood pool, heart walls, and great vessels from MRI scans.
Standard automated methods usually rely on "atlases" (templates of healthy hearts), but CHD anatomy is so variable that these templates often fail. Consequently, clinicians must manually segment 100–200 slices, a process that takes 4–8 hours per patient. This research introduces an interactive tool that leverages expert input efficiently to make 3D modeling a routine clinical reality.
Methodology: Patch-Based Label Propagation
The proposed framework utilizes a Patch-Based Interactive Segmentation approach. Instead of relying on global anatomical models, the algorithm focuses on local appearance.
- User Input: A clinician labels a small number of "reference" regions.
- Label Propagation: The algorithm segments the rest of the volume by analyzing local $5 \times 5$ intensity patches. For every unlabeled voxel, it identifies the $k$ most similar patches in the reference set and uses majority voting to assign a label (blood pool, myocardium, or background).
- Spatial Weighting: To ensure accuracy, the similarity search is weighted by distance. Patches physically closer to the reference labels are given higher priority, helping the algorithm track structures that change shape across the volume.
Active Learning and ROI Targeting
A core innovation of this work is determining which part of the heart the user should label next to maximize information gain. The authors compared two primary active learning strategies using "Oracle" simulations (using ground-truth data to simulate perfect error detection):
- Entire Slices: Querying the user to label a full short-axis plane.
- Regions of Interest (ROIs): Querying the user to label a small $39 \times 39$ pixel box in an area of high uncertainty.
Labeling a specific ROI proved significantly more efficient than labeling an entire slice. In a full slice, much of the user's effort is often wasted on easy-to-segment background or areas the algorithm already handles well. Targeting high-error ROIs provides the most "information gain" per user interaction.
Key Results and Clinical Impact
Validation on high-resolution pediatric MRI scans of DORV patients demonstrated that a high-quality 3D model could be generated using only 14 reference slices (out of approximately 200 total), reducing manual work from several hours to under an hour.
As shown in the figure above, even with just 3 manually labeled slices, the patch-based method can reconstruct the rough global structure of the heart. By 14 slices, the segmentation (indicated by red and yellow contours) accurately captures complex internal structures like the Ventricular Septal Defect (VSD) and the orientation of the Aorta (AO) and Pulmonary Artery (PA).
Evaluating Active Learning Strategies
Quantitative analysis showed that the ROI-based active learning strategy improved the Dice coefficient (a measure of overlap accuracy) for the myocardium by approximately 5 points compared to labeling entire slices, given the same amount of user effort.
Interestingly, the study found that a "uniform distribution" (spacing slices evenly) is a very strong baseline. As seen in the comparison of strategies, random slice selection often leads to initial performance drops, whereas targeted ROI selection consistently drives the highest accuracy improvements.
Conclusion and Future Work
Interactive segmentation offers a safer, more efficient path for clinical use than "black box" AI because it integrates physician expertise into the model-building process. While current computational speeds (approximately one hour per scan) are a limitation, future work using approximate nearest neighbor searches could enable real-time interactivity. Additionally, expanding the tool to allow input on arbitrarily oriented planes could further improve the segmentation of great vessels that run perpendicular to the standard short-axis view.
Summary updated