3D probe localization from 2D ultrasound images using an RFF-enhanced deep neural network.

Area of Science:

• Medical Imaging
• Machine Learning
• Anesthesiology

Background:

• Ultrasound (US) imaging is valuable for peripheral nerve blocking (PNB) due to its non-invasiveness and real-time visualization.
• High costs of 3D US limit its widespread adoption, requiring clinicians to infer 3D anatomy from 2D images.
• Accurate US probe trajectory estimation is crucial for reconstructing 3D volumes from freehand 2D scans.

Purpose of the Study:

• To develop an enhanced tool for estimating the freehand trajectory of an ultrasound probe using 2D US images.
• To improve the generalization capability of deep learning models for 2D US probe localization.
• To facilitate 3D ultrasound volume reconstruction for improved PNB procedures.

Main Methods:

• A kernel-based deep learning approach was employed for US probe trajectory estimation.
• Random Fourier Features (RFF) were integrated to enhance model generalization for 2D US probe localization.
• The model was trained on a public dataset of anatomical phantoms and validated using cross-validation.

Main Results:

• The proposed RFF-based deep learning enhancement significantly improved the accuracy of 2D US probe localization.
• The RFF layer demonstrated superior performance compared to baseline models in predicting probe trajectories.
• Cross-validation confirmed the robustness of the developed method across various training data splits.

Conclusions:

• The RFF-based deep learning method offers a promising solution for accurate freehand ultrasound probe trajectory estimation.
• This advancement can aid in reconstructing 3D ultrasound volumes from 2D data, potentially reducing reliance on expensive 3D US systems.
• The developed tool has the potential to enhance the precision and accessibility of ultrasound-guided procedures like PNB.