An Instance Segmentation Model Based on Deep Learning for Intelligent Diagnosis of Uterine Myomas in MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Women's Health

Background:

• Uterine myomas affect 70% of women, impacting fertility and health.
• Manual MRI analysis for myomas is time-consuming, subjective, and challenging due to complex anatomy.
• Accurate identification of uterine structures and myomas is crucial for clinical treatment planning.

Purpose of the Study:

• To develop an automated instance segmentation network for uterine MRI.
• To accurately identify the location, category, and masks of the uterine wall, uterine cavity, and myomas.
• To provide an objective reference for surgical planning and improve diagnostic efficiency.

Main Methods:

• Designed a novel backbone network to learn diverse shape features and reduce background noise.
• Optimized anchor box generation for improved bounding box prediction and regression.
• Implemented an adaptive iterative subdivision strategy for realistic and accurate mask boundary details.

Main Results:

• The proposed network achieved superior average precision (AP) compared to state-of-the-art models.
• Achieved improvements in AP: 8.8% for uterine wall, 8.4% for uterine cavity, and 3.2% for myomas.
• Demonstrated the first successful multiclass instance segmentation in uterine MRI.

Conclusions:

• The developed network offers a convenient and objective tool for uterine myoma diagnosis.
• Significant value in improving diagnostic efficiency and enabling automatic auxiliary diagnosis.
• Facilitates the clinical development of appropriate surgical plans for uterine myomas.