Abstract
The proposed semi-supervised method is capable of cardiac segmentation and chamber volume quantification in a CHD population with wide anatomical variability. It accurately delineates the heart chambers and great vessels and can be used to accurately calculate ventricular and atrial volumes throughout the cardiac cycle. Such a segmentation method can reduce inter- and intra- observer variability and make CMR exams more standardized and efficient.
Introduction
Congenital heart disease (CHD) remains the leading cause of perinatal death. As medical advancements improve survival rates, a growing adult population requires lifelong cardiac monitoring. Cardiovascular magnetic resonance (CMR) is the gold standard for assessment, but traditional 2D "cine" imaging relies on thick slices and precise, expert-dependent positioning, which introduces measurement variability.
While 3D cine CMR provides high-resolution, isotropic data of the entire heart, it presents a significant data processing challenge. Manually segmenting every structure across the 30 frames of a cardiac cycle is prohibitively time-consuming for clinical routine. This study addresses the need for an automated tool capable of handling the extreme anatomical variability in CHD patients, ranging from standard heart layouts to complex single-ventricle geometries.
Methodology
The researchers developed a pipeline based on a modified 3D U-Net architecture. The primary innovation is a semi-supervised training strategy designed to minimize manual labeling effort while maintaining high accuracy across the temporal dimension of the cardiac cycle.
Label Propagation Strategy
Instead of requiring experts to label all 30 frames, the workflow only requires annotations for the end-diastole (ED) and end-systole (ES) frames—the points where the heart is at its fullest and emptiest. The model then uses these as anchors to propagate labels through the cycle:
- Propagator Network: A model trained on the manual ED and ES frames predicts segmentations for their immediate temporal neighbors.
- Iterative Refinement: These new "pseudo-labels" are added to the training set to predict the next set of frames, iteratively "propagating" labels through the entire 30-frame cycle.
- Weighted-Probabilistic Training: To account for potential error accumulation, higher weights are assigned to frames closer to the original manual ground truth during the final training phase.
Dataset and Evaluation
The study utilized a dataset of 99 patients representing a spectrum of CHD complexity:
- Normal Anatomy (n=26): Standard heart structures with minor issues like arrhythmias.
- Mild/Moderate Complexity (n=39): Repaired defects such as Tetralogy of Fallot.
- Severe Complexity (n=34): Highly abnormal structures, such as single ventricles or transposed great arteries.
The data was partitioned into 74 training, 13 validation, and 12 independent testing cases.
Key Results
The semi-supervised method significantly outperformed standard supervised approaches trained only on the anchor frames:
- Segmentation Accuracy: Achieved a significantly higher Dice score of 83.23%, compared to 77.98% for the standard supervised method.
- Volumetric Precision: Bias in heart chamber volume calculation—a critical metric for surgical planning—was reduced to ≤ 5.2 ml, nearly half the bias of the supervised version (≤ 10.1 ml).
- Clinical Utility: Automated 3D measurements showed strong agreement with manual 2D clinical standards, suggesting the tool is ready for more rigorous clinical validation.
Discussion and Future Work
The study demonstrates that manual labels for just two extreme cardiac phases contain enough information to bootstrap a robust model for the entire cycle. The semi-supervised approach was particularly effective at tracking the subtle, complex motion of the atria compared to standard methods.
However, challenges remain in "severe" CHD cases where anatomy is highly unpredictable. Small or highly mobile structures like the coronary sinus also remain difficult to segment accurately. Future improvements could involve incorporating patient diagnostic metadata directly into the network to guide the segmentation process in cases with highly irregular geometries.
Summary updated