Computerized breast cancer analysis system using three stage semi-supervised learning method.

Area of Science:

• Medical Imaging
• Computer-Aided Detection (CAD)
• Machine Learning

Background:

• Training effective CAD systems requires substantial labeled medical image data, which is time-consuming and ethically complex to acquire.
• Unlabeled data offers a viable solution to overcome limitations in labeled data availability for CAD development.

Purpose of the Study:

• To develop and evaluate a novel three-stage semi-supervised learning (SSL) scheme for enhancing CAD system performance.
• To investigate the impact of incorporating unlabeled data alongside limited labeled data in a CAD system.

Main Methods:

• A three-stage SSL scheme was implemented, integrating data weighing, feature selection, and a co-training data labeling algorithm.
• Global density asymmetry features were utilized to reduce false positives in the CAD system.
• The system's performance was evaluated using mammograms from 400 women, with 90 labeled and the remainder unlabeled regions.

Main Results:

• The proposed SSL scheme achieved a highest Area Under the Curve (AUC) of 0.841, outperforming the use of labeled data alone by 7.4%.
• The performance improvement peaked when approximately 60 labeled data points were used in conjunction with unlabeled data.
• The study demonstrated a significant enhancement in CAD performance through the integration of unlabeled data.

Conclusions:

• The developed three-stage semi-supervised learning approach effectively improves CAD system performance by leveraging unlabeled data.
• The findings suggest that utilizing unlabeled data is a promising strategy for advancing computerized cancer research and future CAD applications.