Deep Learning-Based CT-Less Cardiac Segmentation of PET Images: A Robust Methodology for Multi-Tracer Nuclear...

Area of Science:

• Nuclear Cardiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Quantitative cardiovascular PET/CT imaging is crucial for diagnosing cardiac pathologies.
• Current segmentation methods rely on CT images, which have limitations due to CT-PET mismatch and poor performance on low-dose scans.
• Deep learning (DL) models often struggle with CT-based cardiac segmentation in PET/CT imaging.

Purpose of the Study:

• To develop a DL-based methodology for direct cardiac segmentation from PET images.
• To overcome the limitations of CT-based segmentation in cardiovascular PET/CT imaging.
• To achieve accurate and robust segmentation of the whole heart and its components.

Main Methods:

• Developed a DL-based methodology using the nnU-Net V2 pipeline for direct cardiac PET image segmentation.
• Trained models on 406 cardiac PET images (¹⁸F-FDG, ¹³N-NH₃, ⁸²Rb) from 146 patients.
• Validated the models on an external set of 15 cardiac images, comparing segmentation accuracy using Dice coefficient, Jaccard distance, and surface distance metrics.

Main Results:

• Achieved high average Dice coefficients for whole heart segmentation (0.932 internal, 0.941 external).
• Demonstrated robust performance for segmenting left myocardium (0.88), left ventricle (0.828), and right ventricle (0.876).
• Reported an average volume prediction error of less than 2% for cardiac components, with no significant differences across radiotracers or validation folds.

Conclusions:

• An automated DL pipeline for cardiac PET image segmentation was successfully developed.
• The methodology provides accurate and robust segmentation, outperforming CT-based approaches.
• This technique enhances the reliability of quantitative cardiovascular imaging for diagnosing heart conditions.