Abstract

ABSTRACT: Accurate lesion segmentation is a critical component of computer-aided diagnosis for breast cancer, as it enables precise lesion delineation and robust quantitative assessment. However, breast lesion segmentation remains challenging because of tumor heterogeneity, variations in fibroglandular tissue, and the complex morphology of breast lesions. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has emerged as an effective modality for the early detection and characterization of breast lesions 6 due to its ability to capture detailed information on tumor morphology and microenvironment. This study proposes a two-stage framework for breast DCE-MRI lesion segmentation. In the first stage, bounded turning Mittag-Leffler enhancement is applied to improve image quality and enhance lesion-relevant structures. In the second stage, an extended Visual Geometry Group (VGG)-based network with dilated convolution is employed to segment breast lesions. The proposed framework was evaluated on a public multicenter dataset comprising 979 pre-treatment T1-weighted cases and 10,863 bitmap slice images converted from original Neuroimaging Informatics Technology Initiative (NIFTI) volumes with expert voxel-level annotations. The dataset was divided into training, validation, and testing subsets at ratios of 68%, 12%, and 20%, respectively. Based on 5-fold cross-validation, the proposed method achieved an accuracy of 98.73%, a Dice coefficient of 91.56%, and a Jaccard index of 85.41%. Furthermore, the framework demonstrated competitive performance compared with related studies. These findings confirm the effectiveness of the proposed framework for breast DCE-MRI lesion segmentation and highlight its potential to improve the accuracy and reliability of breast cancer computer-aided diagnosis systems.