Deep Learning-Based Analysis of Mammographic Images for Breast Cancer Detection Using Transfer Learning

초록

Breast cancer ranks as one of the most perilous diseases, being the second most prevalent disease among women globally. This condition involves a rapid and uncontrolled proliferation of breast cells. Prominent symptoms of breast cancer include alterations in breast size and shape, discomfort, the presence of lumps, and nipple discharge. Primarily affecting women aged 50–60 years and older, this malignancy exhibits a substantial occurrence rate within the elderly demographic, accounting for over 80% of cases. Unfortunately, breast cancer claims the lives of numerous women annually. However, early detection and diagnosis offer a chance for a positive outcome, potentially reducing the fatality rate by 20% and enhancing patient survival rates. Past research endeavours have centred on training diverse models to detect breast cancer in its nascent stages. The objective of this research is to detect breast cancer at an early stage and facilitate medical fertility. In this work, we proposed an innovative transfer learning methodology for identifying breast cancer through mammograms. This approach not only aids medical professionals in automating diagnosis but also addresses potential human errors during analysis. By employing transfer learning, our technique empowers doctors to detect cancer in its incipient stages, substantially enhancing diagnostic accuracy. The convolutional neural network (CNN) architectures explored were Inception-v3, ResNet-50, VGG-16, SqueezeNet, and AlexNet. The dataset was collected from the local hospital of Lahore, comprising 900 mammogram images. After segmenting the mammogram images and eliminating noise, we conducted a comprehensive assessment of the CNN architectures. Remarkably, the AlexNet architecture outperformed its counterparts, achieving an outstanding accuracy of 96.7%. This achievement underscores the potency of our transfer learning model combined with distinct solver types. © 2025 Elsevier B.V., All rights reserved.

키워드