Weakly and semi-supervised segmentation of uterus using coarse annotations in MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Endometrial cancer poses significant risks through invasion and metastasis.
• Accurate uterine segmentation on MR images is crucial for assessing myometrial infiltration in endometrial cancer and quantifying uterine volume for benign diseases.
• Current deep learning segmentation methods require time-consuming and subjective manual pixel-level labeling.

Purpose of the Study:

• To develop an efficient uterus segmentation framework for MR images.
• To reduce reliance on manual pixel-level annotations by combining weakly and semi-supervised learning.
• To improve segmentation performance by filtering incorrect pseudo-labels using a confidence-guided strategy.

Main Methods:

• A novel framework combining weakly supervised and semi-supervised learning for uterus segmentation.
• A two-branch network incorporating residual blocks and dual perturbations.
• Implementation of a confidence-guided strategy to refine pseudo-labels by filtering incorrect pixels.

Main Results:

• The proposed method demonstrates superior performance compared to existing weakly and semi-supervised techniques, particularly with limited labeled data.
• Achieved segmentation performance comparable to fully supervised methods across all labeling rates.
• Validated on MR images from 220 patients with various uterine diseases.

Conclusions:

• The developed framework offers an effective solution for uterus segmentation, reducing annotation burden.
• The confidence-guided strategy enhances segmentation accuracy by improving pseudo-label quality.
• This approach provides essential support for clinical disease assessment and preoperative planning in uterine pathologies.