Privileged Modality Distillation for Vessel Border Detection in Intracoronary Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Disease Diagnosis

Background:

• Intracoronary imaging visualizes coronary artery tissue, crucial for diagnosing coronary artery disease.
• Vessel border detection in intracoronary images (VBDI) is vital for computer-aided diagnosis.
• Existing VBDI methods struggle with vessel-environment variability, limiting performance.

Purpose of the Study:

• To propose a novel Privileged Modality Distillation (PMD) framework for robust VBDI.
• To address the limitations of existing VBDI methods caused by data variability.
• To improve the accuracy and adaptability of VBDI across different intracoronary imaging modalities.

Main Methods:

• Developed a Privileged Modality Distillation (PMD) framework to transform single-input-single-task (SIST) to multiple-input-multiple-task (MIMT) learning.
• Utilized a privileged image modality to enhance learning in the target modality.
• Implemented PMD using a structure-deformable neural network, refining MIMT back to SIST for testing.

Main Results:

• The proposed PMD framework demonstrated effectiveness in VBDI, achieving a Dice index greater than 0.95.
• PMD significantly improved model adaptation to diverse vessel environments.
• The method outperformed six state-of-the-art VBDI techniques in experiments.

Conclusions:

• The novel PMD framework successfully addresses vessel-environment variability in VBDI.
• PMD enables effective VBDI across different intracoronary imaging modalities without requiring the privileged modality during testing.
• This approach offers a significant advancement for computer-aided diagnosis in coronary artery disease.