Myocardial perfusion imaging SPECT left ventricle segmentation with graphs

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Computational Anatomy

Background:

• Myocardial perfusion imaging (MPI) using single-photon emission computed tomography (SPECT) relies on various collimators to assess coronary artery disease (CAD).
• The choice of collimator and the incorporation of prior left ventricular (LV) shape information can influence diagnostic outcomes in MPI-SPECT.
• Deep learning (DL) methods are increasingly used for cardiac image segmentation, but their performance can be affected by imaging protocol variations.

Purpose of the Study:

• To evaluate the impact of incorporating prior information into LV segmentation compared to DL approaches.
• To assess the differences in LV geometry segmentation across four distinct collimation techniques (multi-pinhole, LEHR, CardioC, CardioD) on multiple datasets.
• To determine how collimator choice affects the projected cardiac geometry in MPI-SPECT.

Main Methods:

• A novel continuous graph-based approach utilizing the continuous max-flow (CMF) min-cut algorithm was developed for automatic LV segmentation.
• The method incorporated prior knowledge of cardiac geometry.
• Performance was evaluated using precision, recall, Intersection over Union (IoU), and Dice score metrics on 80 patient datasets across different collimator types.

Main Results:

• The developed segmentation method demonstrated superior performance compared to DL approaches, achieving higher scores in most evaluation metrics.
• Receiver operating characteristic (ROC) curve analysis revealed varying stability across different collimators.
• Uniform Manifold Approximation and Projection (UMAP) analysis confirmed that LV shapes projected by different collimators are distinguishable.

Conclusions:

• Incorporating prior information into segmentation significantly enhances the performance of MPI-SPECT analysis.
• Collimation strategies exert a substantial influence on the projected cardiac geometry, impacting diagnostic interpretation.
• The findings suggest that optimizing collimator selection and leveraging prior shape information are crucial for accurate CAD assessment using MPI-SPECT.