Detecting Asymmetric Patterns and Localizing Cancers on Mammograms.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Breast Cancer Diagnostics

Background:

• Mammography is crucial for breast cancer screening, but current computer-aided diagnosis systems often lack simultaneous lesion localization for biopsy.
• Identifying suspicious regions of interest (ROIs) alongside global cancer status prediction remains a challenge.

Purpose of the Study:

• To develop deep learning networks that can simultaneously identify lesions (masses, microcalcifications) and classify breast cancer status from mammograms.
• To investigate the utility of incorporating breast asymmetry information into deep learning models for improved lesion detection and classification.

Main Methods:

• Development of deep learning networks to analyze paired mammograms for masses and microcalcifications.
• Implementation of strategies to exclude false positives and stepwise performance improvement using asymmetric breast information.
• Evaluation of the dual-purpose model in the Digital Mammography DREAM Challenge.

Main Results:

• The developed deep learning method achieved a co-leading performance in the Digital Mammography DREAM Challenge for breast cancer prediction.
• The model successfully identified suspicious regions of interest (ROIs) for potential biopsy.
• Incorporating asymmetric information improved the model's ability to detect lesions and classify cancer status.

Conclusions:

• A dual-purpose deep learning approach can effectively identify suspicious lesions and classify breast cancer status simultaneously.
• Utilizing breast asymmetry in deep learning models enhances diagnostic performance in mammography.
• This method provides valuable localization of potential lesions, aiding in biopsy decisions and improving overall breast cancer diagnosis.