Efficient Pix2Vox++ for 3D Cardiac Reconstruction from 2D echo views.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Cardiovascular Diagnostics

Background:

• Accurate geometric quantification of the human heart is crucial for diagnosing and managing cardiac diseases.
• Ultrasound imaging, the primary cardiac modality, demands high operator skill and is prone to artifacts, complicating analysis.
• Current limitations in 2D ultrasound capabilities hinder comprehensive 3D anatomical assessment.

Purpose of the Study:

• To develop a pipeline for reconstructing 3D cardiac anatomy from limited 2D ultrasound data.
• To optimize existing Pix2Vox++ networks for reduced memory and computational complexity in 3D reconstruction.
• To enable less operator-dependent cardiac imaging analysis and discover new biomarkers.

Main Methods:

• Modification of Pix2Vox++ networks to decrease memory usage and computational load.
• Development of a pipeline for 3D anatomical reconstruction from standard 2D cardiac views.
• Evaluation using synthetically generated data and preliminary testing with real echocardiographic images.

Main Results:

• Achieved accurate 3D whole-heart reconstructions from only two standard 2D cardiac views.
• Demonstrated a peak intersection over union (IoU) score exceeding 0.88 in synthetic data evaluations.
• Presented preliminary successful results using real echocardiographic images.

Conclusions:

• The proposed pipeline effectively enables 3D anatomical reconstruction from limited 2D ultrasound data.
• The optimized network architecture significantly reduces computational requirements for 3D cardiac modeling.
• This approach holds promise for improving cardiac diagnosis and reducing reliance on expert operator skills.