Point Cloud Registration for Fusion between SPECT MPI and CTA Images

1. Problem Statement

The clinical fusion of Single Photon Emission Computed Tomography Myocardial Perfusion Imaging (SPECT MPI) and Computed Tomography Angiography (CTA) is critical for simultaneously assessing myocardial perfusion (functional) and coronary anatomy (structural). However, this fusion faces significant challenges:

• Cross-Modality Misregistration: Differences in imaging physics, resolution, reconstruction algorithms, and patient positioning lead to misalignment.
• Low Resolution & Artifacts: SPECT images suffer from low spatial resolution and motion artifacts, making direct intensity-based registration difficult.
• Reliance on Manual Landmarks: Current methods often rely on manual landmark selection, which is time-consuming, subjective, and prone to error.
• Initialization Failure: Direct fine-registration algorithms often fail due to large initial scale and orientation differences between the two modalities.

2. Methodology

The authors propose a coarse-to-fine registration framework that integrates U-Net segmentation, automated anatomical landmark extraction, scale-space normalization, and comparative fine registration. The pipeline consists of four main stages:

A. Automated Segmentation and Point Cloud Generation

• CTA Data: A U-Net model segments both the Left Ventricle Epicardial Region (LVER) and Right Ventricle Epicardial Region (RVER). Surface point clouds are extracted from these segments.
• SPECT Data: A specialized pipeline resamples short-axis (SAX) images to a long-axis (LAX) perspective to reduce geometric distortion. A pre-trained U-Net segments the LVER, and the resulting surface is mapped back to the original Cartesian coordinate system to generate a point cloud.

B. Automated Landmark Extraction (Asymmetric Strategy)

To establish a robust initial alignment, the system automatically identifies key anatomical landmarks:

• CTA Landmarks: The intersection plane between the Left and Right ventricles is calculated using Principal Component Analysis (PCA) on contact points (derived via dual k-d tree nearest neighbor queries). A Bézier curve is generated along the junction, and 9 points are sampled as landmarks representing the interventricular septum.
• SPECT Landmarks: Since SPECT lacks clear RV definition, the system estimates the center, apex, and base using RANSAC circle fitting. It then constructs a long-axis rotation sampling strategy to find the plane with maximum coverage density, simulating the interventricular junction. 18 points are sampled along a generated Bézier curve.

C. Coarse Registration (Initialization)

Before fine registration, a Scale-Space Consistency preprocessing step is applied:

1. Scale Normalization: Eigenvalues from the covariance matrices of both point clouds are compared to derive an isotropic scale factor, resizing the SPECT point cloud to match the CTA volume.
2. Landmark-Driven Alignment: The pre-sampled landmarks from both modalities are aligned using the Umeyama algorithm to solve for the optimal rigid transformation (rotation and translation), providing a reliable starting position.

D. Fine Registration

Six different point cloud registration algorithms are applied to the Left Ventricle Epicardial Surface (LVES) point clouds to refine the alignment:

1. ICP (Iterative Closest Point)
2. SICP (Scaled ICP)
3. CPD Rigid (Coherent Point Drift - Rigid)
4. CPD Affine
5. CluReg (Cluster-based Registration)
6. FFD (Free-Form Deformation)
7. BCPD++ (Bayesian Coherent Point Drift++)

The transformation matrices derived from these point cloud registrations are then propagated to the original 3D voxel volumes using cubic spline interpolation (for rigid/affine) or nearest-neighbor interpolation (for non-rigid offsets) to achieve final image fusion.

3. Key Contributions

1. Novel Coarse-to-Fine Framework: A unified pipeline that combines anatomical landmark detection, scale-space normalization, and multi-stage refinement to handle cross-modality registration robustly.
2. Automated Asymmetric Landmarking: A method to automatically extract interventricular junction landmarks from CTA (using dual ventricle geometry) and SPECT (using morphological characteristics), eliminating the need for manual intervention.
3. Algorithm-Agnostic Comparative Study: The framework allows for the plug-and-play evaluation of multiple fine-registration algorithms within the same standardized pipeline. This identifies BCPD++ as the superior performer while proving the framework's value lies in the initialization strategy rather than a single algorithm.
4. Clinical Applicability: The approach preserves sub-millimeter coronary details from CTA while accurately overlaying quantitative SPECT perfusion data.

4. Experimental Results

The study was evaluated on a retrospective cohort of 60 patients (66 total CTA cases for training U-Net).

• Segmentation Performance: The U-Net achieved a Dice Similarity Coefficient (DSC) of 0.9531 for the Left Ventricle and 0.8460 for the Right Ventricle.
• Registration Accuracy (with Coarse Initialization):
  • BCPD++ outperformed all other methods (ICP, SICP, CPD, CluReg, FFD).
  • Mean Point Cloud Distance (MPE): 1.7 mm (vs. 3.8 mm for ICP).
  • Apex Displacement Error (AE): 6.7° (vs. 12.0° for ICP).
  • Mean Interventricular Groove Distance (MGE): 8.0 mm.
  • Groove Plane Center Error (GCE): 6.6 mm.
• Importance of Coarse Registration: A control experiment skipping the coarse registration step showed catastrophic failure for most algorithms. For instance, without coarse initialization, the MPE for BCPD++ rose slightly (1.6 mm), but the Apex Displacement Error jumped to 51.1° and MGE to 38.6 mm, proving that fine registration alone cannot correct global orientation and scale errors.

5. Significance

This work provides a practical, automated solution for SPECT-CTA fusion, addressing the critical bottleneck of cross-modality misalignment. By achieving millimeter-level accuracy and sub-degree orientation alignment, the framework enables:

• Precise Ischemia Localization: Accurately mapping perfusion defects to specific coronary artery territories.
• Lesion-Level Functional Assessment: Correlating structural stenosis (CTA) with functional impact (SPECT) in a single coordinate system.
• Robustness: The "scale-space consistency" and landmark-driven initialization make the system resilient to the inherent resolution and contrast differences between SPECT and CTA, offering a reliable basis for clinical decision-making in cardiology.

Code Availability: The authors have made the code publicly available at https://github.com/MIILab-MTU/SPECTMPIAndCTAFusion.