Stacked Denoising Autoencoder on Mammography for Breast Lesion Discrimination

Abstract

One of the primary causes of cancer-related mortality for women worldwide is still breast cancer (BC). Therefore, improving patient survival rates requires early and accurate diagnosis. Early detection has been greatly aided by mammography screening, but accurate lesion classification is still difficult because of low image contrast, dense breast tissue, subtle lesion features, and interpretation variation among radiologists. This paper proposes a deep learning (DL)-based breast lesion classification framework that makes use of a Stacked Denoising Autoencoder (SDAE) to reduce feature redundancy, learn robust discriminative features, and improve the model's generalization and classification performance. The performance of the suggested framework is compared with cutting-edge DL models, such as ResNet-50, DCNN, and CBAM-CNN, using 10-fold cross-validation with a 70%/30% training-testing split on the publicly accessible CBIS-DDSM mammography dataset. According to the experimental results, the suggested SDAE consistently outperforms the benchmark models with 93.69 percent sensitivity, 95.15 percent specificity, 95.74 percent classification accuracy, and 96.58 percent AUC (Area Under the Receiver Operating Characteristic Curve). This suggests that the suggested framework is useful in distinguishing between benign and malignant breast lesions.