DeepCAT: Deep Computer-Aided Triage of Screening Mammography

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Deep learning has advanced breast cancer detection but not mammography triage.
• Efficient triage of mammograms for radiologist review is crucial for timely diagnosis.

Purpose of the Study:

• To develop and evaluate DeepCAT, a deep learning system for mammography triage based on cancer suspicion.
• To assess DeepCAT's capability in discarding non-cancerous images and prioritizing suspicious ones.

Main Methods:

• Developed DeepCAT using 1878 2D mammograms (CC & MLO) from the Digital Database for Screening Mammography.
• The system comprises a mammogram classifier cascade and a mass detector to generate a priority score.
• Evaluated on 595 testing images for triage accuracy and prioritization effectiveness.

Main Results:

• DeepCAT recommended low priority for 53% of images (315/595), none of which contained malignant masses.
• The system's prioritization ordering required an average of 26 adjacent swaps for a perfect review order.
• Demonstrated potential to significantly improve efficiency for breast imagers.

Conclusions:

• DeepCAT shows promise in enhancing mammography workflow efficiency.
• The system can effectively triage mammograms, prioritizing those with potential malignant masses for radiologist review.