Improving annotation efficiency for fully labeling a breast mass segmentation dataset

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Breast cancer is a leading cause of death in women, necessitating effective early detection.
• Screening programs aim for early detection, but a shortage of labeled image datasets hinders AI model development.
• Improving AI models for breast cancer detection requires addressing the bottleneck of creating large, accurately labeled datasets.

Purpose of the Study:

• To present an active learning (AL) framework for segmenting breast masses in 2D digital mammography.
• To reduce the manual effort required from expert annotators in creating labeled mammographic datasets.
• To publish a novel dataset of segmented breast masses to facilitate research in AI-driven early detection.

Main Methods:

• Developed an active learning (AL) framework for pixel-wise binary segmentation of mammographic masses.
• Created a dataset of 1136 mammographic masses with segmentation labels, validated by radiologists.
• Simulated a human annotator within the AL framework to train and evaluate AI-assisted labeling methods.

Main Results:

• Reduced expert annotator input by 44% through the active learning framework.
• Published a dataset of 1136 binary segmentation labels for mammographic masses.
• Validated the quality of proposed labels using intersection over union on a public dataset.

Conclusions:

• Active learning significantly enhances the efficiency and time-effectiveness of creating labeled mammogram datasets.
• The proposed framework aids in developing high-quality datasets with minimized manual effort.
• This research supports advancements in digital mammography and AI-based breast cancer screening.