Large scale deep learning for computer aided detection of mammographic lesions

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Support
• Machine Learning in Radiology

Background:

• Machine learning advances have led to deep neural networks (DNNs) with successful pattern recognition applications.
• Computer-aided detection (CAD) systems in mammography traditionally rely on manually designed features.
• The potential for AI to independently interpret mammograms necessitates comparison with existing methods.

Purpose of the Study:

• To compare a state-of-the-art mammography CAD system with a Convolutional Neural Network (CNN).
• To evaluate the performance of a CNN for independent mammogram interpretation.
• To investigate the added value of traditional features and contextual information for CNN performance.

Main Methods:

• A Convolutional Neural Network (CNN) was trained and compared against a traditional CAD system.
• Both systems were trained on a large dataset of approximately 45,000 mammograms.
• A reader study compared the CNN's performance against certified screening radiologists on a patch level.

Main Results:

• The CNN outperformed the traditional CAD system at low sensitivity and performed comparably at high sensitivity.
• Incorporating manual features, location, and patient information improved CNN performance, especially at high specificity.
• The reader study found no significant difference between the CNN and certified radiologists in patch-level analysis.

Conclusions:

• CNNs demonstrate strong potential for mammogram analysis, outperforming traditional CAD systems in key areas.
• Hybrid approaches combining CNNs with contextual features can further enhance diagnostic accuracy.
• The developed CNN shows performance comparable to human experts, suggesting its utility in breast cancer screening.