Improving Workflow Efficiency for Mammography Using Machine Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Mammography screening is crucial for early breast cancer detection.
• Radiologists face high workloads, necessitating efficient screening processes.
• Accurate identification of normal mammograms can optimize radiologist time.

Purpose of the Study:

• To evaluate machine learning's ability to reduce radiologist workload by identifying normal mammograms.
• To assess if AI can accurately classify mammograms, prioritizing uncertain and abnormal cases for review.

Main Methods:

• Utilized a dataset of over 7,000 mammograms from UK National Health Service Breast Screening Program centers.
• Employed a convolutional neural network (CNN) with multitask learning to extract imaging and non-imaging features.
• A deep neural network fused multiple mammogram views for diagnosis and assessment recommendation.

Main Results:

• The model achieved a negative predictive value of 0.99.
• Identified 34% of negative mammograms in a 15% cancer prevalence test set.
• Identified 91% of negative mammograms in a 1% cancer prevalence test set.

Conclusions:

• Machine learning successfully reduced the workload of normal mammogram interpretation for radiologists.
• The AI model maintained diagnostic accuracy while decreasing the number of mammograms requiring radiologist review.