Implicit neural representation for scalable 3D reconstruction from sparse ultrasound images

Area of Science:

• Medical imaging
• Computational anatomy
• Neural networks

Background:

• Volumetric ultrasound imaging is valuable but limited by 2D array costs and data reconstruction challenges.
• Existing 3D reconstruction algorithms struggle with scale and require precise transducer positioning, which is often unavailable.
• Current methods often rely on discrete interpolation, limiting real-time analysis and accuracy.

Purpose of the Study:

• To develop a novel, efficient, and accurate method for reconstructing 3D ultrasound volumes from 2D slices.
• To overcome the limitations of existing discrete reconstruction algorithms and transducer position dependency.
• To enable continuous and scalable 3D volume reconstruction using implicit neural representations.

Main Methods:

• Utilized a 1D ultrasound array on a programmable motor for data acquisition.
• Developed and employed implicit neural representations for continuous 3D volume reconstruction.
• Compared the proposed neural network approach against classic reconstruction algorithms.

Main Results:

• The implicit neural network achieved 7.9x performance improvement over classic algorithms while maintaining accuracy.
• A reconstruction pipeline demonstrated 93% accuracy on simulated data using only 36 B-mode images.
• In vivo evaluation in mice showed a 6.3% mean error in measuring tumor volumes.

Conclusions:

• Implicit neural representations offer a viable solution for reconstructing 3D ultrasound volumes from limited 2D data.
• The proposed method significantly reduces the data requirements and computational complexity compared to traditional interpolation techniques.
• This approach facilitates interactive analysis and has potential applications in preclinical and clinical volumetric ultrasound imaging.