Convolutional Neural Networks for Early Diagnosing of Breast Cancer

Abstract

Due to the high demand for mammograms, radiologists are swamped with many patients' mammograms. Radiologists' workload has increased dramatically over the last 15 years during on-call hours. This rise is due to an increase in the number of computed tomography (CT). Previous research published in the early 2000s found a 22% increase in radiology examinations during on-call hours in the United States over four years. The shortage of radiologists has recently worsened due to the COVID-19 pandemic. Based on their training and experience, radiologists classify mammography images as benign or malignant. Automating this diagnostic process with a machine learning algorithm may improve diagnosis speed and accuracy. This research aims to propose an effective automated algorithm system that lessens the workload of radiologists while improving speed and accuracy. This study uses Convolution Neural Networks (CNN) to assist radiologists in classifying lesions by incorporating repeated imaging. In addition, longitudinal electronic health record (EHR) data prior to diagnosis will augment the ML algorithm to improve its accuracy. This study takes a novel approach to the fusion model because only a few models have been developed to integrate both clinical and imaging data for diagnostic purposes, and few of them have included both EHR and longitudinal imaging data. Therefore, this study compared various multi-modal fusion and single models that can use pixel-based data (image) and EHR data from a large urban medical center.