BRAIxDet: Learning to detect malignant breast lesion with incomplete annotations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Screening mammograms require accurate malignant lesion detection for early diagnosis.
• Current methods often rely on fully annotated datasets, which are costly and time-consuming to create.
• Real-world datasets frequently contain a mix of fully and weakly annotated images, posing a data utilization challenge.

Purpose of the Study:

• To address the dilemma of utilizing incomplete mammogram annotation datasets for malignant lesion detection.
• To propose a cost-effective and accurate solution for training detection models with mixed annotation types.
• To develop a weakly- and semi-supervised learning framework for malignant breast lesion detection.

Main Methods:

• A two-stage approach: (1) weak supervision pre-training of a multi-view mammogram classifier using the entire dataset.
• (2) Semi-supervised student-teacher learning to extend the classifier into a multi-view detector using both fully and weakly annotated data.

Main Results:

• The proposed method achieves state-of-the-art results on two real-world screening mammogram datasets with incomplete annotations.
• Demonstrates effective utilization of both fully and weakly annotated data for improved detection accuracy.
• Outperforms existing methods in malignant breast lesion detection under incomplete annotation scenarios.

Conclusions:

• The developed weakly- and semi-supervised learning approach offers a practical solution for training malignant breast lesion detection models with incomplete annotations.
• This method enhances the utility of large, real-world mammogram datasets, potentially improving screening efficiency and accuracy.
• The findings suggest a significant advancement in automated detection of breast cancer from screening mammograms.