Abstract
This paper presents a novel 4D statistical shape model for the segmentation of the left myocardium in gated SPECT image sequences. By extending the implicit level set representation to include time as a fourth dimension, the method leverages temporal correlations in cardiac motion to overcome the limitations of low-resolution and high-noise SPECT data. Principal Component Analysis (PCA) is applied to a training set of 4D volumes to create a shape prior that constrains the segmentation process. Experimental results demonstrate that this 4D approach significantly outperforms traditional 3D methods, reducing average relative symmetric voxel error from 53% to 28%.
Introduction
Accurate segmentation of the left ventricle (LV) is essential for assessing cardiac function through clinical metrics such as ejection fraction. However, Single Photon Emission Computed Tomography (SPECT) imaging presents significant challenges due to low spatial resolution and high noise levels (low signal-to-noise ratio).
Traditional segmentation techniques often treat cardiac sequences as independent 3D volumes, failing to exploit the predictable temporal patterns of heart movement. Furthermore, explicit 4D models based on landmarks or meshes suffer from the "nightmare" of establishing point-to-point correspondences across patients and time steps—a task that is both error-prone and computationally difficult in higher dimensions. This work proposes an implicit 4D level set approach to address these issues.
Methodology
The proposed method utilizes an implicit representation where the heart surface is defined as the zero-level set of a 4D function $\phi(x, y, z, t)$.
- 4D Statistical Shape Prior: A training set of 15 hand-segmented gated perfusion SPECT recordings was used to build a shape model. PCA was performed directly on these 4D level set functions, capturing both spatial morphology and temporal dynamics (contraction patterns) without requiring manual point correspondences.
- Energy Minimization: Segmentation is achieved by minimizing an energy functional consisting of a data-driven term (based on the Chan-Vese model for region intensities) and a shape-driven term. The shape term penalizes deviations from the learned 4D shape space, ensuring anatomically plausible results even in noisy frames.
- Transformation Handling: The model incorporates parameters for spatial and temporal translation and scaling to ensure proper alignment between the prior and the target image sequence.
Results and Discussion
The 4D method was evaluated against a state-of-the-art 3D PCA-based baseline (Tsai et al.). The 3D approach frequently resulted in over-segmentation, particularly at end-diastole, with an average relative symmetric voxel error of 53% (ranging from 35% to 90%).
In contrast, the 4D method provided more stable and anatomically plausible results, achieving a significantly lower average error of 28% (ranging from 20% to 40%). The inclusion of temporal information allowed the model to "correct" noisy or ambiguous frames by relying on the learned motion patterns from the rest of the sequence. Even when the test subject was not part of the training set, the 4D prior provided a much more specific constraint than the 3D version.
Conclusion
By treating time as a fourth coordinate within an implicit level set framework, this research demonstrates a robust method for cardiac segmentation in challenging imaging modalities. The 4D shape prior effectively captures the spatio-temporal correlations of the myocardium, offering superior performance over 3D-only models and avoiding the complexities of explicit landmark matching. Future work will focus on expanding the training dataset and incorporating pathological motion patterns into the statistical prior.
Summary updated