Automatic 3-D Lamina Curve Extraction From Freehand 3-D Ultrasound Data Using Sequential Localization Recurrent...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spine Biomechanics

Background:

• Freehand 3-D ultrasound is a noninvasive, affordable tool for spine exams.
• Accurate extraction of 3-D lamina landmarks is crucial for spine shape analysis.
• Challenges include contrast variations, artifacts, and probe manipulation in ultrasound sequences.

Purpose of the Study:

• To develop and evaluate a novel deep learning model, sequential localization recurrent convolutional networks (SL-RCNs), for precise 3-D lamina curve extraction from ultrasound data.
• To enhance the accuracy of 3-D spine shape analysis using freehand ultrasound imaging.

Main Methods:

• Proposed sequential localization recurrent convolutional networks (SL-RCNs) model incorporating contextual relationships and transformation matrix features.
• Analyzed 3-D ultrasound sequences from ten healthy participants (lumbar and thoracic regions).
• Evaluated performance using sevenfold cross-validation with a leave-one-participant-out strategy and tested on three participants.

Main Results:

• SL-RCN reduced mean distance errors (MDEs) from 1.62/1.63 mm to 1.41/1.40 mm.
• Normalized discrete Fréchet distance (NDFD) improved from 0.5910/0.6389 to 0.4276/0.4567.
• Minimal increase (<0.05) in mean NDFD between cross-validation and test data, indicating robust performance.

Conclusions:

• SL-RCN effectively extracts accurate, paired, smooth lamina landmark curves from freehand 3-D ultrasound sequences.
• The model demonstrates significant improvements over previous 2-D analysis methods.
• SL-RCN shows potential for advancing 3-D spine imaging analysis in clinical settings.