DeepHeartCT: A fully automatic artificial intelligence hybrid framework based on convolutional neural network and...

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medical Imaging
• Deep Learning for Medical Image Analysis

Background:

• Cardiac computed tomography angiography (CTA) is crucial for cardiovascular assessment.
• Deep learning (DL) shows promise in medical image segmentation but requires extensive labeled data.
• Manual labeling for DL training in cardiac CTA is labor-intensive and time-consuming.

Purpose of the Study:

• To develop DeepHeartCT, a fully automatic AI system for accurate and rapid cardiac CTA segmentation.
• To leverage DL with computer-generated labels to overcome manual annotation limitations.
• To validate the system's performance on a large clinical dataset.

Main Methods:

• Developed DeepHeartCT, an AI system using deep convolutional neural networks (CNNs).
• Trained the system on a large clinical dataset (n=1,064) using computer-generated labels from a multi-atlas AI system.
• Employed a reverse ranking strategy to select optimal computer-generated labels for training.
• Validated the model on an independent dataset (n=60) with manual labels, using Dice score, Hausdorff distance, and mean surface distance.

Main Results:

• Achieved high segmentation accuracy with a median Dice score of 0.90.
• Demonstrated low median Hausdorff distance (7 mm) and mean surface distance (0.80 mm).
• The framework performed well even when trained on a smaller optimal dataset (n=110), with reduced training time.
• The system showed generalizability for large-scale medical imaging applications.

Conclusions:

• DeepHeartCT provides accurate and rapid segmentation of cardiac CTA.
• The AI system effectively overcomes the challenge of limited labeled data in DL training.
• The framework is robust and suitable for large-scale clinical applications in cardiovascular imaging.