Semi-Supervised Breast Lesion Segmentation Using Confidence-Ranked Features and Bi-Level Prototypes.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Automated breast lesion segmentation in ultrasound (BUS) images is crucial for computer-aided diagnosis.
• Challenges include laborious data annotation, ambiguous lesion boundaries, and low contrast in BUS images.

Purpose of the Study:

• To propose a novel semi-supervised breast segmentation framework, CoBiNet, to overcome current segmentation challenges.
• To enhance the accuracy and efficiency of automated breast lesion segmentation.

Main Methods:

• Developed a semi-supervised framework (CoBiNet) utilizing confidence-ranked features and bi-level prototypes.
• Employed a dual-branch architecture (classifier and projector) with multilevel feature ranking.
• Implemented trans-confidence level (TCL) contrastive learning and cross-guidance (CG) consistency learning.

Main Results:

• CoBiNet demonstrated superior performance over state-of-the-art methods on BUSI and UDIAT datasets.
• The framework effectively handles ambiguous boundaries and low contrast in BUS images.
• Improved recognition of ambiguous regions through feature confidence ranking and contrastive learning.

Conclusions:

• CoBiNet offers a robust and effective solution for semi-supervised breast ultrasound image segmentation.
• The proposed methods significantly advance computer-aided diagnosis in breast imaging.
• Future work will involve releasing the code for broader research application.