Pose Estimation of Ultrasound Probe Using CNN and RNN with Image Reconstruction Loss.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate 3D ultrasound (US) image reconstruction from 1D-array probe sequences is challenging.
• Precise estimation of the probe's pose (position and orientation) is critical for high-fidelity 3D US imaging.
• Conventional methods often struggle to achieve the required accuracy for complex US scans.

Purpose of the Study:

• To develop a novel method for accurate probe pose estimation using Convolutional Neural Networks (CNN).
• To train the CNN model using an image reconstruction loss function for improved pose accuracy.
• To validate the proposed method's performance against existing techniques in 3D US image reconstruction.

Main Methods:

• A Convolutional Neural Network (CNN) architecture was designed for probe pose estimation.
• The CNN was trained using an image reconstruction loss, calculated by a dedicated network.
• The reconstruction network employed an encoder-decoder structure to generate intermediate US images.

Main Results:

• The proposed CNN-based method demonstrated efficient probe pose estimation capabilities.
• Experimental results showed superior performance compared to conventional probe pose estimation methods.
• The image reconstruction loss effectively guided the CNN training for accurate pose determination.

Conclusions:

• The developed CNN method provides an effective solution for accurate probe pose estimation in 3D ultrasound.
• Training with image reconstruction loss is a viable strategy for enhancing pose estimation accuracy.
• This approach offers a promising advancement for high-quality 3D US image reconstruction using 1D-array probes.