Comparison between Deep learning architectures to improve breast cancer detection on screening mammography

Abstract

Recently, developing deep learning based diagnosing systems for breast cancer has gained considerable attention from the research community. It has been proven that employing deep learning techniques in the medical sector is an effective way to speed up generating results of diagnosing as well as contributing to early detection of cases. Developing accurate deep learning based systems is tightly-coupled with effective architectures, which in turn means taking into consideration the type and number of layers is critical in this context. This work proposes a deep learning based model for detecting breast cancer from mammogram medical images. Four diagnosing models are developed using Resnet from scratch, ResNet50 V2, VGG16, and MobileNet V2. The intelligent models are trained using the RSNA Screening Mammography Breast Cancer Detection dataset available on the Kaggle website. In the pre-processing step, EightSymmetry and cutout based augmentation techniques are employed for the purpose of enhancing accuracy. The models are tested without augmentation, with EightSymmetry based augmentation, and with cutout based augmentation. The results showed that using cutout augmentation with the VGG16-based model performs the best in terms of AUC (0.8) when compared to other models, where Resnet from scratch (0.6), ResNet50 V2 (0.7), and MobileNet V2 (0.79).