Deep learning-based plane pose regression in obstetric ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics

Background:

• Acquiring 3D fetal anatomical understanding from 2D ultrasound images is challenging for trainees.
• Current methods often require additional sensors for precise spatial orientation.

Purpose of the Study:

• To develop an ultrasound plane localization system for 3D fetal visualization and training.
• To achieve this without integrating external sensors into the scanning process.

Main Methods:

• A regression convolutional neural network (CNN) was designed to estimate the 6D pose of ultrasound planes.
• The network was trained on synthetic 3D ultrasound volumes and fine-tuned on real fetal scans.
• Training data included planes sliced at various orientations, with emphasis on the transventricular plane.

Main Results:

• Phantom data showed median errors of <1.21 degrees and <1 mm for plane localization.
• Real data demonstrated successful generalization with median errors of 11.84 mm and 25.17 degrees.
• The average inference time per plane was 2.97 ms.

Conclusions:

• The CNN reliably localizes ultrasound planes in the fetal brain using phantom data.
• The system shows promise for generalizing to unseen fetal brains of similar gestational age.
• Future work aims to extend this to whole-fetus visualization and freehand ultrasound navigation.