Spatio-Temporal Partitioning and Description of Full-Length Routine Fetal Anomaly Ultrasound Scans.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Clinical Workflow Optimization

Background:

• Routine fetal anomaly ultrasound scans generate extensive video data.
• Manual analysis of these scans is time-consuming and subjective.
• Automating clinical workflow description can improve efficiency and consistency.

Purpose of the Study:

• To develop and evaluate deep learning models for automatic description of fetal anomaly ultrasound scan workflows.
• To investigate the effectiveness of spatio-temporal video analysis techniques.
• To assess the correlation between automated and manual workflow characterization.

Main Methods:

• Investigated multiple deep learning architectures: 2D CNN, 2D+t CNN, LSTM, and convolutional LSTM.
• Analyzed contributions of short-term and long-term temporal changes in video data.
• Utilized a multi-stream framework for spatio-temporal video analysis.

Main Results:

• The multi-stream framework achieved the highest performance with top-1 accuracy of 0.77 and top-3 accuracy of 0.94.
• Automated partitioning and characterization of unlabeled scans showed high correlation (ρ=0.95, p=0.0004) with manual labels.
• Deep learning models effectively captured spatio-temporal dynamics in ultrasound videos.

Conclusions:

• Deep learning, particularly the multi-stream framework, offers a practical solution for automatic clinical workflow description of fetal ultrasound scans.
• The proposed methods demonstrate high accuracy and strong correlation with manual annotations.
• Automated analysis has the potential to significantly streamline and standardize obstetric ultrasound procedures.