Deep-learning based multibeat echocardiographic cardiac phase detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics

Background:

• Automatic cardiac phase detection in multibeat echocardiograms is vital for clinical measurements.
• Current methods are limited by high data annotation costs and slow processing times.

Purpose of the Study:

• Introduce EchoPhaseNet, a novel deep learning network for fast, accurate cardiac phase detection.
• Address limitations of low annotation costs and limited data for variable-length echocardiographic sequences.

Main Methods:

• Utilized five echocardiographic datasets (Echo-DT, PhaseDetection, EchoNet-Dynamic, CAMUS, EchoNet-Dynamic-MultiBeat).
• Trained and validated EchoPhaseNet against four other deep learning methods.
• Evaluated performance using GradCAM for visualization and absolute frame difference (aFD) for accuracy, with statistical significance testing.

Main Results:

• EchoPhaseNet achieved effective phase detection using only ED/ES labels, reducing annotation costs.
• Demonstrated superior or comparable accuracy (aFD) to existing methods on internal and external datasets.
• Exhibited significantly faster inference times (under 8 ms on RTX 4080 GPU) compared to other methods.

Conclusions:

• EchoPhaseNet offers significant advantages in reduced annotation costs and increased detection speed.
• The model shows strong generalization capabilities across diverse datasets.
• Presents a practical and promising solution for clinical multibeat echocardiographic cardiac phase detection.