Adaptive Density-Guided Multi-Scale CLAHE combined with Edge-Attention U-Net for Dense Breast Visualization and Segmentation

Abstract

Breast density is an important factor in the interpretation of mammograms as dense fibroglandular tissue can make it more difficult to see the edges of lesions. Contrast enhancement and deep learning-based segmentation methods have been much researched, but current methods often use fixed enhancement parameters and fail to maintain tissue structure in cases of heterogeneous dense breast. While standard segmentation networks can have less sensitivity towards dense tissue boundaries, conventional Contrast Limited Adaptive Histogram Equalization (CLAHE) can add noise and cause over-enhancement. This inspiring the development of a density-aware enhancement and segmentation framework.  In this study, they introduce a new Adaptive Density-Guided Multi-Scale CLAHE (ADG-MSCLAHE) combined with Edge-Attention U-Net architecture for dense breast visualization and segmentation. Mammographic density parameters are first estimated adaptively based on mammographic density characteristics to determine enhancement parameters. The multi scale CLAHE outputs at different tile sizes are merged using density dependent weighting coefficients to enhance anatomical structures along with contrast. These feature maps are then combined in an edge-aware manner before being fed into an Attention U-Net for a better delineation of dense tissues and boundary preservation. The methodology was tested on the normal, benign and malignant dense breast cases in the combined INbreast–DDSM–MIAS mammographic dataset. Experimental results show that, the proposed enhancement framework provides PSNR value as 31.24 dB, SSIM value as 0.846 and Contrast Improvement Index (CII) as 1.91 which is better when compared with conventional CLAHE and fixed multi-scale enhancement. The proposed Edge-Attention U-Net achieved better dice coefficient score of 0.921, IoU score of 0.854 and F1 score of 0.921 for segmentation which showed a better localization of dense tissue. The results show that density-guided enhancement and edge-aware segmentation enhances the visibility of tissues, retains structural information and produces clinically interpretable dense breast visualization. The proposed scheme is an effective computational framework for mammographic analysis and computer-aided applications for breast screening.