Automatic Segmentation and Alignment of Uterine Shapes from 3D Ultrasound Data

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Reproductive Medicine

Background:

• Uterine shape is crucial for fertility and pregnancy outcomes.
• Existing uterine shape classifications lack real-world measurements and large datasets.
• 3D ultrasound advances offer improved gynecological healthcare exploration.

Purpose of the Study:

• To establish the normal uterine shape using real-world 3D vaginal ultrasound scans.
• To fill the gap in large-scale uterine shape studies.
• To facilitate research into uterine shape abnormalities linked to infertility and recurrent miscarriages.

Main Methods:

• Developed an automated system for uterine shape segmentation and alignment from 3D ultrasound data.
• Employed deep learning for automatic uterine segmentation.
• Utilized standard geometrical approaches for shape alignment.

Main Results:

• The automated system achieved high accuracy in segmenting and aligning uterine shapes.
• Segmentation achieved an average Dice Similarity Coefficient (DSC) of 0.90.
• The alignment method showed minimal translation and rotation errors, yielding a preliminary average shape consistent with expert findings.

Conclusions:

• An automated approach for uterine shape segmentation and alignment from 3D ultrasound data was presented.
• A deep learning model (nnU-Net) achieved high accuracy.
• A publicly available dataset of 3D transvaginal ultrasound volumes with manual annotations was created to support further research.