Searching collaborative agents for multi-plane localization in 3D ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Three-dimensional (3D) ultrasound (US) offers superior spatial and diagnostic information compared to 2D US.
• 3D US acquisition can capture multiple standard planes (SPs) simultaneously, improving efficiency.
• Manual localization of SPs in 3D US is difficult due to low image quality, large search spaces, and anatomical variations.

Purpose of the Study:

• To develop an automated framework for simultaneous multi-standard plane localization in 3D US.
• To enhance user-independence and scanning efficiency in 3D US examinations.
• To address the challenges of manual SP localization in complex 3D US datasets.

Main Methods:

• A novel multi-agent reinforcement learning (MARL) framework was proposed for simultaneous SP localization.
• A recurrent neural network (RNN) based collaborative module was integrated to improve inter-agent communication and spatial relationship learning.
• Neural architecture search (NAS) was employed to automatically optimize the network architecture for agents and the collaborative module.

Main Results:

• The proposed MARL framework achieved high accuracy in localizing multiple SPs across challenging 3D US datasets (uterus and fetal brain).
• Average localization accuracies of 7.03°/1.59mm (uterus) and 9.75°/1.19mm (fetal brain) were reported.
• The lightweight MARL model demonstrated superior accuracy compared to existing state-of-the-art methods.

Conclusions:

• The developed MARL framework offers an effective and accurate solution for automatic SP localization in 3D US.
• This approach is generalizable to different challenging US datasets and can handle anatomical variations, including normal and abnormal uterus cases.
• The method represents a significant advancement, being the first to achieve automatic SP localization in pelvic US volumes.