Lesion detection model from mammography images using deep learning

Abstract

[ background ] Breast cancer screening has been performed with mammography images. Mammography reading skills are known to be varied among readers. Artificial intelligence (AI) could help to increase overall accuracy rate and to reduce work overload to make the system more efficient. Previous study revealed the promising result on automatic interpretation of mammography images. Our aim is to clarify whether the accuracy rate interpreting various types of mammography is not inferior to the previous study. [materials and methods] KBOG, the non-profit organization in Kanagawa prefecture has initiated the clinical trial, called KBOG1701 trial. Nine institutes had prepared more than a thousand mammography images with annotation information. Cropped images were created from mammography images and were used to train AI. [result] Based on cropped images, the accuracy, sensitivity, specificity and AUC rates were 81%, 88%, 74%, 0.880, respectively. Trained model was implemented on a web server to automatically scan mammography images on the internet. [conclusion] Using CAD system based on AI could be helpful to reduce false negative risk and increase overall accuracy rate. AI could be the game changer on the breast screening system.