Neural architecture search of echocardiography view classifiers

Area of Science:

• Medical imaging analysis
• Artificial intelligence in cardiology
• Ultrasound technology

Background:

• Echocardiography is a primary clinical tool for cardiac assessment.
• Accurate identification of probe viewpoints is crucial for automated echocardiographic analysis.
• Current methods may lack efficiency or require extensive data.

Purpose of the Study:

• To develop automated methods for identifying 14 anatomical echocardiographic views.
• To design efficient neural network architectures for rapid and accurate cardiac view classification.
• To investigate factors influencing model performance, including image quality and dataset size.

Main Methods:

• Utilized convolutional neural networks (CNNs) for automated echocardiographic view identification.
• Employed a differentiable architecture search approach to create compact and fast neural networks.
• Evaluated models on a dataset of 8732 videos from 374 patients, analyzing 14 view classes.

Main Results:

• Proposed models achieved high classification performance (accuracy 88.4%-96%, precision 87.8%-95.2%, recall 87.1%-95.1%).
• Models demonstrated significantly reduced trainable parameters (up to 99.9% reduction) compared to deeper architectures.
• Real-time inference times per image ranged from 3.6 to 12.6 ms, indicating high efficiency.

Conclusions:

• The developed CNN models are faster and achieve comparable performance to standard architectures for echocardiographic view classification.
• These models require less training data, making them more accessible.
• The proposed approach enables real-time detection of standard echocardiographic views, aiding clinical practice.