Multibeat echocardiographic phase detection using deep neural networks.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate identification of end-diastolic and end-systolic frames in echocardiography is crucial but challenging.
• Manual selection is subjective, impacting critical measurements like myocardial strain.
• Automated frame detection is highly desirable for improved accuracy and efficiency.

Purpose of the Study:

• To develop and validate a deep neural network for automatic detection of end-diastolic and end-systolic frames.
• To assess the model's performance against expert annotations and inter-observer variability.
• To provide a robust automated solution for echocardiographic analysis.

Main Methods:

• Deep neural networks were developed and trained on multi-center patient echocardiographic data.
• The model was tested on apical four-chamber 2D multibeat cine loops of varying lengths.
• Expert cardiologists provided ground-truth annotations for performance comparison.

Main Results:

• The model achieved average frame differences of -0.09 ± 1.10 for end-diastolic and 0.11 ± 1.29 for end-systolic frames compared to ground-truth.
• Testing on an unseen clinical site yielded average differences of -1.34 ± 3.27 and -0.31 ± 3.37 frames.
• All detection errors were within the range of inter-observer variability.

Conclusions:

• The automated model accurately identifies multiple end-diastolic and end-systolic frames in echocardiographic videos.
• Its performance is indistinguishable from human experts.
• The model offers significantly reduced processing time compared to manual methods.