Predicting malignancy from mammography findings and image-guided core biopsies

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Mammography is crucial for breast cancer screening.
• Accurate prediction of mass density and malignancy improves diagnostic accuracy.
• Machine learning offers potential for automated analysis of mammographic findings.

Purpose of the Study:

• To develop machine learning models for predicting breast mass density and malignancy from mammography.
• To evaluate the performance of various algorithms in classifying mammographic findings.
• To assess the utility of a mass density predictor in improving malignancy prediction.

Main Methods:

• Utilized a dataset of 348 breast masses from 328 female subjects.
• Applied and varied parameters of multiple machine learning algorithms, including Support Vector Machines (SVM).
• Trained models to predict mass density and malignancy based on annotated mammography findings.

Main Results:

• An SVM-based model achieved 81.3% accuracy in predicting mass density, outperforming expert annotation (70%).
• An SVM-based model achieved 85.6% accuracy and 85% positive predictive value for malignancy prediction.
• The developed model can predict malignancy even when mass density data is missing, by utilizing the density predictor.

Conclusions:

• Machine learning, particularly SVMs, can effectively predict breast mass density and malignancy from mammography.
• The integrated approach, using a density predictor to impute missing data, enhances malignancy prediction capabilities.
• This study demonstrates the potential of AI in improving the efficiency and accuracy of mammographic interpretation for breast cancer diagnosis.