Abstract:
Currently, the identification of any anomaly in mammography images presents a lot of challenges,
counting the complexity of data patterns, the determination of suitable deep learning algorithms,
and the requirement for computational efficiency. To overcome these obstacles, it is pivotal to
create a strong deep learning algorithm with viable training protocols. This research centers on
progressing both effectiveness and feature extraction methods for breast cancer detection from
mammograms. It addresses the existing challenges while also investigating methodologies to
improve algorithms performance. The objective of this framework is to execute an effective
strategy for the abnormality in (mediolateral oblique (MLO) and cranial-caudal (CC) views) of
9,752 mammography images by utilizing augmentation methods. To boost diagnostic results and
speed, we propose a novel computer-aided detection (CAD) framework that leverages an advanced
deep learning system, which includes VGG16, EfficientNetB7, and DenseNet121. Our approach
incorporates ensemble learning with convolutional neural network (CNN) models and applies
transfer learning in all three architectures. The results illustrate the impressive accuracy of the
proposed methods as 82.61%, 84.75%, 91.49%, and 88.89% for the VGG16, EfficientNetB7,
DenseNet121, and ensemble models, respectively. This inventive framework marks a critical step
forward in the early location of breast cancer, giving the potential for more precise and convenient
analysis.
