Abstract
Automated analysis of 4D cardiac MRI is a critical yet challenging task in clinical cardiology. This paper presents a fully automated pipeline for the segmentation and tracking of the left ventricle (LV), right ventricle (RV), and myocardium throughout the cardiac cycle. By employing non-rigid registration based on free-form deformations (FFDs) and a dual-atlas strategy, the method achieves high-fidelity tracking that correlates strongly with manual expert segmentations. The approach effectively addresses common MRI challenges such as low tissue contrast and complex anatomical motion.
Introduction
Quantitative cardiac analysis—calculating metrics like ejection fraction, wall motion, and chamber volumes—is essential for diagnosing and managing heart disease. However, manual segmentation of 4D MRI datasets is a labor-intensive "clinician’s nightmare," prone to subjective bias and inter-observer variability.
Automating this process is difficult due to several factors: the fuzzy borders between the myocardium and surrounding tissues, significant anatomical variation between patients, and the complex, non-rigid deformation (twisting and squeezing) the heart undergoes during a single beat. This research proposes a robust solution by treating segmentation as a mapping problem rather than a simple boundary-finding task.
Methodology
The core of the proposed system is Non-rigid Registration utilizing Free-Form Deformations (FFDs). The transformation model operates in two stages: a global affine transformation for basic alignment (scaling and rotation), followed by a local deformation model based on B-splines that warps the image to match fine anatomical details.
Registration and Similarity
To ensure accurate warping, the system uses Normalised Mutual Information (NMI) as a similarity measure. NMI is particularly effective for MRI because it remains robust even when intensity levels between the atlas and the patient scan do not match perfectly.
The Two-Atlas Strategy
- Population-Specific Atlas: A "gold standard" 3D model derived from manual segmentations of 14 healthy subjects.
- Subject-Specific Atlas: To mitigate noise in individual scans, the authors register all time frames of a patient's cycle to the first frame and average them, creating a high-contrast, low-noise "average" image for that specific patient.
Tracking via Propagation
Instead of segmenting every frame independently, the algorithm segments the initial end-diastolic frame and then "propagates" these labels through the sequence. This is achieved by calculating the motion fields (transformations) between consecutive time frames and applying them to the initial segmentation.
Experimental Results
The method was validated using 141 individual 3D volumes from 9 healthy volunteers. Automated volumes were compared against manual expert tracings using linear regression analysis.
The results demonstrated exceptional accuracy, particularly in tracking:
- Left Ventricle (LV): $r = 0.99$
- Myocardium: $r = 0.98$
- Right Ventricle (RV): $r = 0.96$
While tracking accuracy was near-perfect, initial segmentation accuracy was also high (LV: 0.94, RV: 0.92, Myocardium: 0.83), confirming that the non-rigid registration framework is highly capable of following the heart's complex motion once the initial anatomy is identified.
Discussion and Conclusion
This research highlights that temporal correlation is key to robust cardiac analysis. By treating the heart's motion as a predictable deformation field, the system achieves results that rival human experts. A significant innovation is the use of subject-specific averaging to improve the signal-to-noise ratio, making the final segmentation more reliable.
However, limitations remain. The system's performance is heavily dependent on the accuracy of the initial end-diastolic frame segmentation. Furthermore, while the results are excellent for healthy volunteers, future work must address pathological hearts (e.g., those with infarcts or hypertrophy) where motion patterns may deviate significantly from the population atlas. The authors suggest that these temporal transformations could eventually be used to measure regional strain, providing a powerful tool for detecting localized ischemia.
Summary updated