AAWS-Net: Anatomy-aware weakly-supervised learning network for breast mass segmentation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Accurate breast mass segmentation is crucial for computer-aided diagnosis of breast cancer.
• Limited annotated data and under-utilization of weak annotations hinder deep learning model generalization.
• Acquiring high-quality annotations is resource-intensive in medical imaging.

Purpose of the Study:

• To develop an efficient and accurate breast mass segmentation method using weak annotations.
• To leverage anatomy features from mammograms for improved model performance.
• To address the data scarcity challenge in deep learning for medical image analysis.

Main Methods:

• Proposed an Anatomy-Aware Weakly-Supervised learning Network (AAWS-Net) based on teacher-student architectures.
• Employed a weakly-supervised strategy in the teacher model for anatomy structure extraction via image reconstruction.
• Utilized knowledge distillation for morphological difference learning and integrated prior knowledge into the student network.

Main Results:

• AAWS-Net demonstrated promising performance in breast mass segmentation on the CBIS-DDSM dataset.
• Achieved competitive segmentation accuracy and Intersection over Union (IoU) compared to state-of-the-art methods.
• Effectively extracted useful information from mammograms with weak annotations.

Conclusions:

• The proposed AAWS-Net effectively utilizes weak annotations for accurate breast mass segmentation.
• The method enhances the localization and segmentation capabilities of deep learning models.
• AAWS-Net offers a viable solution for data-scarce scenarios in medical image analysis.