An Integrated Multimodal-Based CAD System for Breast Cancer Diagnosis.

Abstract

Simple Summary: Diagnosis of breast cancer goes through multiple processes. Recently, a variety of system-aided diagnosis (CAD) systems have been proposed as Primary systems for initial diagnosis based on mammogram screenings. This paper enhances the diagnosis accuracy by using mammograms of both sides of the patient’s breasts instead of the infected side only. In addition, the paper boosts CAD accuracy by adding patient information and medical history along with mammogram images’ features. The proposed multimodal approach will serve as the nucleus for future work at both data and system levels to diagnose breast cancer and other diseases caused by various factors.

Abstract: Breast cancer has been one of the main causes of death among women recently, and it has been the focus of attention of many specialists and researchers in the health field. Because of its seriousness and spread speed, breast cancer-resisting methods, early diagnosis, diagnosis, and treatment have been the points of research discussion. Many computers-aided diagnosis (CAD) systems have been proposed to reduce the load on physicians and increase the accuracy of breast tumor diagnosis. To the best of our knowledge, combining patient information, including medical history, breast density, age, and other factors, with mammogram features from both breasts in craniocaudal (CC) and mediolateral oblique (MLO) views has not been previously investigated for breast tumor classification. In this paper, we investigated the effectiveness of using those inputs by comparing two combination approaches. The soft voting approach, produced from statistical information-based models (decision tree, random forest, K-nearest neighbor, Gaussian naive Bayes, gradient boosting, and MLP) and an image-based model (CNN), achieved 90% accuracy. Additionally, concatenating statistical and image-based features in a deep learning model achieved 93% accuracy. We found that it produced promising results that would enhance the CAD systems. As a result, this study finds that using both sides of mammograms outperformed the result of using only the infected side. In addition, integrating the mammogram features with statistical information enhanced the accuracy of the tumor classification. Our findings, based on a novel dataset, incorporate both patient information and four-view mammogram images, covering multiple classes: normal, benign, and malignant.