Deep learning-based classification of coronary arteries and left ventricle using multimodal data for autonomous...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Optimal X-ray imaging parameters are vital for coronary angiography and structural cardiac procedures, balancing image quality and radiation dose.
• Manual selection of anatomy-dependent imaging parameters increases clinical workload and complexity.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for autonomous detection of cardiac anatomies (left coronary artery, right coronary artery, left ventricle) from single X-ray frames.
• To enable automatic adjustment of imaging parameters by selecting appropriate organ programs based on detected anatomy.

Main Methods:

• Compared three deep learning architectures (ResNet-50, MLP, multimodal) using a dataset of 275 clinical radiographic sequences (coronary angiography, left ventriculography).
• Utilized the first non-contrast frame for anatomy detection, allowing pre-contrast system adaptation.
• Evaluated models on an independent test set of 146 sequences.

Main Results:

• The multimodal deep learning model achieved an average F1 score of 0.82 and an AUC of 0.87, matching expert performance.
• The model accurately classified cardiac anatomies from pre-contrast frames without visible structures.
• Demonstrated effective prediction of cardiac anatomy for cine acquisitions.

Conclusions:

• The proposed deep learning model accurately predicts cardiac anatomy, facilitating automatic selection of imaging parameters.
• This AI-driven approach has the potential to optimize image quality, reduce radiation exposure, streamline workflows, and improve patient safety in cardiac procedures.