Weakly Supervised Breast Ultrasound Image Segmentation Based on Image Selection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate segmentation of Breast Ultrasound (BUS) images is crucial for computer-aided diagnosis systems.
• Fully supervised methods require challenging pixel-level annotations, limiting their practical application.
• Weakly supervised learning offers a promising alternative to reduce annotation burden.

Purpose of the Study:

• To develop a novel weakly supervised learning framework for BUS image segmentation.
• To address the challenge of ineffective samples in existing weakly supervised methods.
• To improve the robustness and accuracy of BUS image segmentation models.

Main Methods:

• A novel image selection technique using Class Activation Maps (CAMs) to identify high-quality training samples.
• Utilization of the 'Segment Anything' model for efficient pseudo-segmentation label generation.
• Training a segmentation model via the Mean Teacher method, incorporating both pseudo-labeled and unlabeled data.

Main Results:

• The proposed framework demonstrates a significant improvement in BUS image segmentation.
• Achieved an Intersection over Union (IoU) score that is 82.9% of fully supervised methods.
• Successfully reduced the dependency on extensive manual annotations.

Conclusions:

• The developed weakly supervised learning approach is effective for BUS image segmentation.
• The image selection and pseudo-labeling strategy enhances model training.
• This method offers a viable path towards more accessible and accurate computer-aided diagnosis for BUS imaging.