SpaceTime-SonoNet: efficient classification of ultra-sound video sequences.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Ultrasound (US) imaging is a vital diagnostic tool.
• Analyzing spatio-temporal information in US sequences can enhance diagnostic accuracy.
• Current AI models may not fully leverage temporal dynamics in US data.

Purpose of the Study:

• To extend the SonoNet architecture for spatio-temporal analysis of ultrasound video sequences.
• To introduce computationally efficient 3D and (2+1)D convolutional neural network variants.
• To evaluate the effectiveness of these models for scan-plane detection and potential AI-driven scan assistants.

Main Methods:

• Development of 3D-SonoNet32 by extending 2D convolutions to 3D.
• Implementation of an efficient (2+1)D SonoNet variant to balance performance and computational cost.
• Experimental validation on a scan-plane detection task, analyzing spatio-temporal benefits.

Main Results:

• The proposed 3D and (2+1)D SonoNet architectures effectively capture spatio-temporal information from ultrasound sequences.
• Space-time models demonstrate significant benefits over purely spatial models for ultrasound analysis.
• The models are computationally lightweight and achieve competitive performance, suitable for real-time deployment.

Conclusions:

• Novel space-time SonoNet architectures offer enhanced analysis of ultrasound video data.
• These models can improve the quality and reproducibility of ultrasound examinations.
• The developed methodologies support the development of AI-driven scan assistants for clinical use.